text
stringlengths 307
2.02k
| inputs
dict | prediction
null | prediction_agent
null | annotation
stringclasses 2
values | annotation_agent
stringclasses 1
value | vectors
null | multi_label
bool 1
class | explanation
null | id
stringlengths 36
36
| metadata
null | status
stringclasses 2
values | metrics
dict | label
class label 2
classes |
---|---|---|---|---|---|---|---|---|---|---|---|---|---|
TITLE: WikiLingua: A New Benchmark Dataset for Cross-Lingual Abstractive Summarization
ABSTRACT: We introduce WikiLingua, a large-scale, multilingual dataset for the
evaluation of crosslingual abstractive summarization systems. We extract
article and summary pairs in 18 languages from WikiHow, a high quality,
collaborative resource of how-to guides on a diverse set of topics written by
human authors. We create gold-standard article-summary alignments across
languages by aligning the images that are used to describe each how-to step in
an article. As a set of baselines for further studies, we evaluate the
performance of existing cross-lingual abstractive summarization methods on our
dataset. We further propose a method for direct crosslingual summarization
(i.e., without requiring translation at inference time) by leveraging synthetic
data and Neural Machine Translation as a pre-training step. Our method
significantly outperforms the baseline approaches, while being more cost
efficient during inference. | {
"abstract": "We introduce WikiLingua, a large-scale, multilingual dataset for the\nevaluation of crosslingual abstractive summarization systems. We extract\narticle and summary pairs in 18 languages from WikiHow, a high quality,\ncollaborative resource of how-to guides on a diverse set of topics written by\nhuman authors. We create gold-standard article-summary alignments across\nlanguages by aligning the images that are used to describe each how-to step in\nan article. As a set of baselines for further studies, we evaluate the\nperformance of existing cross-lingual abstractive summarization methods on our\ndataset. We further propose a method for direct crosslingual summarization\n(i.e., without requiring translation at inference time) by leveraging synthetic\ndata and Neural Machine Translation as a pre-training step. Our method\nsignificantly outperforms the baseline approaches, while being more cost\nefficient during inference.",
"title": "WikiLingua: A New Benchmark Dataset for Cross-Lingual Abstractive Summarization",
"url": "http://arxiv.org/abs/2010.03093v1"
} | null | null | new_dataset | admin | null | false | null | 2b06bf83-4565-40c7-bd94-55d327c90489 | null | Validated | {
"text_length": 1034
} | 0new_dataset
|
TITLE: Machine Learning Workflow to Explain Black-box Models for Early Alzheimer's Disease Classification Evaluated for Multiple Datasets
ABSTRACT: Purpose: Hard-to-interpret Black-box Machine Learning (ML) were often used
for early Alzheimer's Disease (AD) detection.
Methods: To interpret eXtreme Gradient Boosting (XGBoost), Random Forest
(RF), and Support Vector Machine (SVM) black-box models a workflow based on
Shapley values was developed. All models were trained on the Alzheimer's
Disease Neuroimaging Initiative (ADNI) dataset and evaluated for an independent
ADNI test set, as well as the external Australian Imaging and Lifestyle
flagship study of Ageing (AIBL), and Open Access Series of Imaging Studies
(OASIS) datasets. Shapley values were compared to intuitively interpretable
Decision Trees (DTs), and Logistic Regression (LR), as well as natural and
permutation feature importances. To avoid the reduction of the explanation
validity caused by correlated features, forward selection and aspect
consolidation were implemented.
Results: Some black-box models outperformed DTs and LR. The forward-selected
features correspond to brain areas previously associated with AD. Shapley
values identified biologically plausible associations with moderate to strong
correlations with feature importances. The most important RF features to
predict AD conversion were the volume of the amygdalae, and a cognitive test
score. Good cognitive test performances and large brain volumes decreased the
AD risk. The models trained using cognitive test scores significantly
outperformed brain volumetric models ($p<0.05$). Cognitive Normal (CN) vs. AD
models were successfully transferred to external datasets.
Conclusion: In comparison to previous work, improved performances for ADNI
and AIBL were achieved for CN vs. Mild Cognitive Impairment (MCI)
classification using brain volumes. The Shapley values and the feature
importances showed moderate to strong correlations. | {
"abstract": "Purpose: Hard-to-interpret Black-box Machine Learning (ML) were often used\nfor early Alzheimer's Disease (AD) detection.\n Methods: To interpret eXtreme Gradient Boosting (XGBoost), Random Forest\n(RF), and Support Vector Machine (SVM) black-box models a workflow based on\nShapley values was developed. All models were trained on the Alzheimer's\nDisease Neuroimaging Initiative (ADNI) dataset and evaluated for an independent\nADNI test set, as well as the external Australian Imaging and Lifestyle\nflagship study of Ageing (AIBL), and Open Access Series of Imaging Studies\n(OASIS) datasets. Shapley values were compared to intuitively interpretable\nDecision Trees (DTs), and Logistic Regression (LR), as well as natural and\npermutation feature importances. To avoid the reduction of the explanation\nvalidity caused by correlated features, forward selection and aspect\nconsolidation were implemented.\n Results: Some black-box models outperformed DTs and LR. The forward-selected\nfeatures correspond to brain areas previously associated with AD. Shapley\nvalues identified biologically plausible associations with moderate to strong\ncorrelations with feature importances. The most important RF features to\npredict AD conversion were the volume of the amygdalae, and a cognitive test\nscore. Good cognitive test performances and large brain volumes decreased the\nAD risk. The models trained using cognitive test scores significantly\noutperformed brain volumetric models ($p<0.05$). Cognitive Normal (CN) vs. AD\nmodels were successfully transferred to external datasets.\n Conclusion: In comparison to previous work, improved performances for ADNI\nand AIBL were achieved for CN vs. Mild Cognitive Impairment (MCI)\nclassification using brain volumes. The Shapley values and the feature\nimportances showed moderate to strong correlations.",
"title": "Machine Learning Workflow to Explain Black-box Models for Early Alzheimer's Disease Classification Evaluated for Multiple Datasets",
"url": "http://arxiv.org/abs/2205.05907v2"
} | null | null | no_new_dataset | admin | null | false | null | 36f5d66c-94a0-4d1c-88c0-f8cd4806b0b1 | null | Validated | {
"text_length": 1995
} | 1no_new_dataset
|
TITLE: A Case Study on the Classification of Lost Circulation Events During Drilling using Machine Learning Techniques on an Imbalanced Large Dataset
ABSTRACT: This study presents machine learning models that forecast and categorize lost
circulation severity preemptively using a large class imbalanced drilling
dataset. We demonstrate reproducible core techniques involved in tackling a
large drilling engineering challenge utilizing easily interpretable machine
learning approaches.
We utilized a 65,000+ records data with class imbalance problem from Azadegan
oilfield formations in Iran. Eleven of the dataset's seventeen parameters are
chosen to be used in the classification of five lost circulation events. To
generate classification models, we used six basic machine learning algorithms
and four ensemble learning methods. Linear Discriminant Analysis (LDA),
Logistic Regression (LR), Support Vector Machines (SVM), Classification and
Regression Trees (CART), k-Nearest Neighbors (KNN), and Gaussian Naive Bayes
(GNB) are the six fundamental techniques. We also used bagging and boosting
ensemble learning techniques in the investigation of solutions for improved
predicting performance. The performance of these algorithms is measured using
four metrics: accuracy, precision, recall, and F1-score. The F1-score weighted
to represent the data imbalance is chosen as the preferred evaluation
criterion.
The CART model was found to be the best in class for identifying drilling
fluid circulation loss events with an average weighted F1-score of 0.9904 and
standard deviation of 0.0015. Upon application of ensemble learning techniques,
a Random Forest ensemble of decision trees showed the best predictive
performance. It identified and classified lost circulation events with a
perfect weighted F1-score of 1.0. Using Permutation Feature Importance (PFI),
the measured depth was found to be the most influential factor in accurately
recognizing lost circulation events while drilling. | {
"abstract": "This study presents machine learning models that forecast and categorize lost\ncirculation severity preemptively using a large class imbalanced drilling\ndataset. We demonstrate reproducible core techniques involved in tackling a\nlarge drilling engineering challenge utilizing easily interpretable machine\nlearning approaches.\n We utilized a 65,000+ records data with class imbalance problem from Azadegan\noilfield formations in Iran. Eleven of the dataset's seventeen parameters are\nchosen to be used in the classification of five lost circulation events. To\ngenerate classification models, we used six basic machine learning algorithms\nand four ensemble learning methods. Linear Discriminant Analysis (LDA),\nLogistic Regression (LR), Support Vector Machines (SVM), Classification and\nRegression Trees (CART), k-Nearest Neighbors (KNN), and Gaussian Naive Bayes\n(GNB) are the six fundamental techniques. We also used bagging and boosting\nensemble learning techniques in the investigation of solutions for improved\npredicting performance. The performance of these algorithms is measured using\nfour metrics: accuracy, precision, recall, and F1-score. The F1-score weighted\nto represent the data imbalance is chosen as the preferred evaluation\ncriterion.\n The CART model was found to be the best in class for identifying drilling\nfluid circulation loss events with an average weighted F1-score of 0.9904 and\nstandard deviation of 0.0015. Upon application of ensemble learning techniques,\na Random Forest ensemble of decision trees showed the best predictive\nperformance. It identified and classified lost circulation events with a\nperfect weighted F1-score of 1.0. Using Permutation Feature Importance (PFI),\nthe measured depth was found to be the most influential factor in accurately\nrecognizing lost circulation events while drilling.",
"title": "A Case Study on the Classification of Lost Circulation Events During Drilling using Machine Learning Techniques on an Imbalanced Large Dataset",
"url": "http://arxiv.org/abs/2209.01607v2"
} | null | null | no_new_dataset | admin | null | false | null | aa7fbe27-3a54-4e61-9172-64185ae18634 | null | Validated | {
"text_length": 2012
} | 1no_new_dataset
|
TITLE: Presenting a Larger Up-to-date Movie Dataset and Investigating the Effects of Pre-released Attributes on Gross Revenue
ABSTRACT: Movie-making has become one of the most costly and risky endeavors in the
entertainment industry. Continuous change in the preference of the audience
makes it harder to predict what kind of movie will be financially successful at
the box office. So, it is no wonder that cautious, intelligent stakeholders and
large production houses will always want to know the probable revenue that will
be generated by a movie before making an investment. Researchers have been
working on finding an optimal strategy to help investors in making the right
decisions. But the lack of a large, up-to-date dataset makes their work harder.
In this work, we introduce an up-to-date, richer, and larger dataset that we
have prepared by scraping IMDb for researchers and data analysts to work with.
The compiled dataset contains the summery data of 7.5 million titles and detail
information of more than 200K movies. Additionally, we perform different
statistical analysis approaches on our dataset to find out how a movie's
revenue is affected by different pre-released attributes such as budget,
runtime, release month, content rating, genre etc. In our analysis, we have
found that having a star cast/director has a positive impact on generated
revenue. We introduce a novel approach for calculating the star power of a
movie. Based on our analysis we select a set of attributes as features and
train different machine learning algorithms to predict a movie's expected
revenue. Based on generated revenue, we classified the movies in 10 categories
and achieved a one-class-away accuracy rate of almost 60% (bingo accuracy of
30%). All the generated datasets and analysis codes are available online. We
also made the source codes of our scraper bots public, so that researchers
interested in extending this work can easily modify these bots as they need and
prepare their own up-to-date datasets. | {
"abstract": "Movie-making has become one of the most costly and risky endeavors in the\nentertainment industry. Continuous change in the preference of the audience\nmakes it harder to predict what kind of movie will be financially successful at\nthe box office. So, it is no wonder that cautious, intelligent stakeholders and\nlarge production houses will always want to know the probable revenue that will\nbe generated by a movie before making an investment. Researchers have been\nworking on finding an optimal strategy to help investors in making the right\ndecisions. But the lack of a large, up-to-date dataset makes their work harder.\nIn this work, we introduce an up-to-date, richer, and larger dataset that we\nhave prepared by scraping IMDb for researchers and data analysts to work with.\nThe compiled dataset contains the summery data of 7.5 million titles and detail\ninformation of more than 200K movies. Additionally, we perform different\nstatistical analysis approaches on our dataset to find out how a movie's\nrevenue is affected by different pre-released attributes such as budget,\nruntime, release month, content rating, genre etc. In our analysis, we have\nfound that having a star cast/director has a positive impact on generated\nrevenue. We introduce a novel approach for calculating the star power of a\nmovie. Based on our analysis we select a set of attributes as features and\ntrain different machine learning algorithms to predict a movie's expected\nrevenue. Based on generated revenue, we classified the movies in 10 categories\nand achieved a one-class-away accuracy rate of almost 60% (bingo accuracy of\n30%). All the generated datasets and analysis codes are available online. We\nalso made the source codes of our scraper bots public, so that researchers\ninterested in extending this work can easily modify these bots as they need and\nprepare their own up-to-date datasets.",
"title": "Presenting a Larger Up-to-date Movie Dataset and Investigating the Effects of Pre-released Attributes on Gross Revenue",
"url": "http://arxiv.org/abs/2110.07039v2"
} | null | null | new_dataset | admin | null | false | null | 14366fed-8b51-40fd-9cc3-c8ff049dd855 | null | Validated | {
"text_length": 2030
} | 0new_dataset
|
TITLE: Human-Machine Collaboration Approaches to Build a Dialogue Dataset for Hate Speech Countering
ABSTRACT: Fighting online hate speech is a challenge that is usually addressed using
Natural Language Processing via automatic detection and removal of hate
content. Besides this approach, counter narratives have emerged as an effective
tool employed by NGOs to respond to online hate on social media platforms. For
this reason, Natural Language Generation is currently being studied as a way to
automatize counter narrative writing. However, the existing resources necessary
to train NLG models are limited to 2-turn interactions (a hate speech and a
counter narrative as response), while in real life, interactions can consist of
multiple turns. In this paper, we present a hybrid approach for dialogical data
collection, which combines the intervention of human expert annotators over
machine generated dialogues obtained using 19 different configurations. The
result of this work is DIALOCONAN, the first dataset comprising over 3000
fictitious multi-turn dialogues between a hater and an NGO operator, covering 6
targets of hate. | {
"abstract": "Fighting online hate speech is a challenge that is usually addressed using\nNatural Language Processing via automatic detection and removal of hate\ncontent. Besides this approach, counter narratives have emerged as an effective\ntool employed by NGOs to respond to online hate on social media platforms. For\nthis reason, Natural Language Generation is currently being studied as a way to\nautomatize counter narrative writing. However, the existing resources necessary\nto train NLG models are limited to 2-turn interactions (a hate speech and a\ncounter narrative as response), while in real life, interactions can consist of\nmultiple turns. In this paper, we present a hybrid approach for dialogical data\ncollection, which combines the intervention of human expert annotators over\nmachine generated dialogues obtained using 19 different configurations. The\nresult of this work is DIALOCONAN, the first dataset comprising over 3000\nfictitious multi-turn dialogues between a hater and an NGO operator, covering 6\ntargets of hate.",
"title": "Human-Machine Collaboration Approaches to Build a Dialogue Dataset for Hate Speech Countering",
"url": "http://arxiv.org/abs/2211.03433v1"
} | null | null | new_dataset | admin | null | false | null | 7846bbb9-9c43-48f8-8277-1d711ec791c7 | null | Validated | {
"text_length": 1152
} | 0new_dataset
|
TITLE: MultiCoNER: A Large-scale Multilingual dataset for Complex Named Entity Recognition
ABSTRACT: We present MultiCoNER, a large multilingual dataset for Named Entity
Recognition that covers 3 domains (Wiki sentences, questions, and search
queries) across 11 languages, as well as multilingual and code-mixing subsets.
This dataset is designed to represent contemporary challenges in NER, including
low-context scenarios (short and uncased text), syntactically complex entities
like movie titles, and long-tail entity distributions. The 26M token dataset is
compiled from public resources using techniques such as heuristic-based
sentence sampling, template extraction and slotting, and machine translation.
We applied two NER models on our dataset: a baseline XLM-RoBERTa model, and a
state-of-the-art GEMNET model that leverages gazetteers. The baseline achieves
moderate performance (macro-F1=54%), highlighting the difficulty of our data.
GEMNET, which uses gazetteers, improvement significantly (average improvement
of macro-F1=+30%). MultiCoNER poses challenges even for large pre-trained
language models, and we believe that it can help further research in building
robust NER systems. MultiCoNER is publicly available at
https://registry.opendata.aws/multiconer/ and we hope that this resource will
help advance research in various aspects of NER. | {
"abstract": "We present MultiCoNER, a large multilingual dataset for Named Entity\nRecognition that covers 3 domains (Wiki sentences, questions, and search\nqueries) across 11 languages, as well as multilingual and code-mixing subsets.\nThis dataset is designed to represent contemporary challenges in NER, including\nlow-context scenarios (short and uncased text), syntactically complex entities\nlike movie titles, and long-tail entity distributions. The 26M token dataset is\ncompiled from public resources using techniques such as heuristic-based\nsentence sampling, template extraction and slotting, and machine translation.\nWe applied two NER models on our dataset: a baseline XLM-RoBERTa model, and a\nstate-of-the-art GEMNET model that leverages gazetteers. The baseline achieves\nmoderate performance (macro-F1=54%), highlighting the difficulty of our data.\nGEMNET, which uses gazetteers, improvement significantly (average improvement\nof macro-F1=+30%). MultiCoNER poses challenges even for large pre-trained\nlanguage models, and we believe that it can help further research in building\nrobust NER systems. MultiCoNER is publicly available at\nhttps://registry.opendata.aws/multiconer/ and we hope that this resource will\nhelp advance research in various aspects of NER.",
"title": "MultiCoNER: A Large-scale Multilingual dataset for Complex Named Entity Recognition",
"url": "http://arxiv.org/abs/2208.14536v1"
} | null | null | new_dataset | admin | null | false | null | 56aa19a9-b936-46d9-9501-40cba4082c0e | null | Validated | {
"text_length": 1375
} | 0new_dataset
|
TITLE: Healthsheet: Development of a Transparency Artifact for Health Datasets
ABSTRACT: Machine learning (ML) approaches have demonstrated promising results in a
wide range of healthcare applications. Data plays a crucial role in developing
ML-based healthcare systems that directly affect people's lives. Many of the
ethical issues surrounding the use of ML in healthcare stem from structural
inequalities underlying the way we collect, use, and handle data. Developing
guidelines to improve documentation practices regarding the creation, use, and
maintenance of ML healthcare datasets is therefore of critical importance. In
this work, we introduce Healthsheet, a contextualized adaptation of the
original datasheet questionnaire ~\cite{gebru2018datasheets} for
health-specific applications. Through a series of semi-structured interviews,
we adapt the datasheets for healthcare data documentation. As part of the
Healthsheet development process and to understand the obstacles researchers
face in creating datasheets, we worked with three publicly-available healthcare
datasets as our case studies, each with different types of structured data:
Electronic health Records (EHR), clinical trial study data, and
smartphone-based performance outcome measures. Our findings from the
interviewee study and case studies show 1) that datasheets should be
contextualized for healthcare, 2) that despite incentives to adopt
accountability practices such as datasheets, there is a lack of consistency in
the broader use of these practices 3) how the ML for health community views
datasheets and particularly \textit{Healthsheets} as diagnostic tool to surface
the limitations and strength of datasets and 4) the relative importance of
different fields in the datasheet to healthcare concerns. | {
"abstract": "Machine learning (ML) approaches have demonstrated promising results in a\nwide range of healthcare applications. Data plays a crucial role in developing\nML-based healthcare systems that directly affect people's lives. Many of the\nethical issues surrounding the use of ML in healthcare stem from structural\ninequalities underlying the way we collect, use, and handle data. Developing\nguidelines to improve documentation practices regarding the creation, use, and\nmaintenance of ML healthcare datasets is therefore of critical importance. In\nthis work, we introduce Healthsheet, a contextualized adaptation of the\noriginal datasheet questionnaire ~\\cite{gebru2018datasheets} for\nhealth-specific applications. Through a series of semi-structured interviews,\nwe adapt the datasheets for healthcare data documentation. As part of the\nHealthsheet development process and to understand the obstacles researchers\nface in creating datasheets, we worked with three publicly-available healthcare\ndatasets as our case studies, each with different types of structured data:\nElectronic health Records (EHR), clinical trial study data, and\nsmartphone-based performance outcome measures. Our findings from the\ninterviewee study and case studies show 1) that datasheets should be\ncontextualized for healthcare, 2) that despite incentives to adopt\naccountability practices such as datasheets, there is a lack of consistency in\nthe broader use of these practices 3) how the ML for health community views\ndatasheets and particularly \\textit{Healthsheets} as diagnostic tool to surface\nthe limitations and strength of datasets and 4) the relative importance of\ndifferent fields in the datasheet to healthcare concerns.",
"title": "Healthsheet: Development of a Transparency Artifact for Health Datasets",
"url": "http://arxiv.org/abs/2202.13028v1"
} | null | null | no_new_dataset | admin | null | false | null | 11460e75-a3e1-47e4-b475-05f92c8cf564 | null | Validated | {
"text_length": 1803
} | 1no_new_dataset
|
TITLE: CoAP-DoS: An IoT Network Intrusion Dataset
ABSTRACT: The need for secure Internet of Things (IoT) devices is growing as IoT
devices are becoming more integrated into vital networks. Many systems rely on
these devices to remain available and provide reliable service. Denial of
service attacks against IoT devices are a real threat due to the fact these low
power devices are very susceptible to denial-of-service attacks. Machine
learning enabled network intrusion detection systems are effective at
identifying new threats, but they require a large amount of data to work well.
There are many network traffic data sets but very few that focus on IoT network
traffic. Within the IoT network data sets there is a lack of CoAP denial of
service data. We propose a novel data set covering this gap. We develop a new
data set by collecting network traffic from real CoAP denial of service attacks
and compare the data on multiple different machine learning classifiers. We
show that the data set is effective on many classifiers. | {
"abstract": "The need for secure Internet of Things (IoT) devices is growing as IoT\ndevices are becoming more integrated into vital networks. Many systems rely on\nthese devices to remain available and provide reliable service. Denial of\nservice attacks against IoT devices are a real threat due to the fact these low\npower devices are very susceptible to denial-of-service attacks. Machine\nlearning enabled network intrusion detection systems are effective at\nidentifying new threats, but they require a large amount of data to work well.\nThere are many network traffic data sets but very few that focus on IoT network\ntraffic. Within the IoT network data sets there is a lack of CoAP denial of\nservice data. We propose a novel data set covering this gap. We develop a new\ndata set by collecting network traffic from real CoAP denial of service attacks\nand compare the data on multiple different machine learning classifiers. We\nshow that the data set is effective on many classifiers.",
"title": "CoAP-DoS: An IoT Network Intrusion Dataset",
"url": "http://arxiv.org/abs/2206.14341v1"
} | null | null | new_dataset | admin | null | false | null | 96fb32b7-5a67-43e2-bf98-3001963116fe | null | Validated | {
"text_length": 1049
} | 0new_dataset
|
TITLE: Multi-feature Dataset for Windows PE Malware Classification
ABSTRACT: This paper describes a multi-feature dataset for training machine learning
classifiers for detecting malicious Windows Portable Executable (PE) files. The
dataset includes four feature sets from 18,551 binary samples belonging to five
malware families including Spyware, Ransomware, Downloader, Backdoor and
Generic Malware. The feature sets include the list of DLLs and their functions,
values of different fields of PE Header and Sections. First, we explain the
data collection and creation phase and then we explain how did we label the
samples in it using VirusTotal's services. Finally, we explore the dataset to
describe how this dataset can benefit the researchers for static malware
analysis. The dataset is made public in the hope that it will help inspire
machine learning research for malware detection. | {
"abstract": "This paper describes a multi-feature dataset for training machine learning\nclassifiers for detecting malicious Windows Portable Executable (PE) files. The\ndataset includes four feature sets from 18,551 binary samples belonging to five\nmalware families including Spyware, Ransomware, Downloader, Backdoor and\nGeneric Malware. The feature sets include the list of DLLs and their functions,\nvalues of different fields of PE Header and Sections. First, we explain the\ndata collection and creation phase and then we explain how did we label the\nsamples in it using VirusTotal's services. Finally, we explore the dataset to\ndescribe how this dataset can benefit the researchers for static malware\nanalysis. The dataset is made public in the hope that it will help inspire\nmachine learning research for malware detection.",
"title": "Multi-feature Dataset for Windows PE Malware Classification",
"url": "http://arxiv.org/abs/2210.16285v1"
} | null | null | new_dataset | admin | null | false | null | 59ca26fc-a4b8-4f78-94dd-16d31625ab97 | null | Validated | {
"text_length": 908
} | 0new_dataset
|
TITLE: Avast-CTU Public CAPE Dataset
ABSTRACT: There is a limited amount of publicly available data to support research in
malware analysis technology. Particularly, there are virtually no publicly
available datasets generated from rich sandboxes such as Cuckoo/CAPE. The
benefit of using dynamic sandboxes is the realistic simulation of file
execution in the target machine and obtaining a log of such execution. The
machine can be infected by malware hence there is a good chance of capturing
the malicious behavior in the execution logs, thus allowing researchers to
study such behavior in detail. Although the subsequent analysis of log
information is extensively covered in industrial cybersecurity backends, to our
knowledge there has been only limited effort invested in academia to advance
such log analysis capabilities using cutting edge techniques. We make this
sample dataset available to support designing new machine learning methods for
malware detection, especially for automatic detection of generic malicious
behavior. The dataset has been collected in cooperation between Avast Software
and Czech Technical University - AI Center (AIC). | {
"abstract": "There is a limited amount of publicly available data to support research in\nmalware analysis technology. Particularly, there are virtually no publicly\navailable datasets generated from rich sandboxes such as Cuckoo/CAPE. The\nbenefit of using dynamic sandboxes is the realistic simulation of file\nexecution in the target machine and obtaining a log of such execution. The\nmachine can be infected by malware hence there is a good chance of capturing\nthe malicious behavior in the execution logs, thus allowing researchers to\nstudy such behavior in detail. Although the subsequent analysis of log\ninformation is extensively covered in industrial cybersecurity backends, to our\nknowledge there has been only limited effort invested in academia to advance\nsuch log analysis capabilities using cutting edge techniques. We make this\nsample dataset available to support designing new machine learning methods for\nmalware detection, especially for automatic detection of generic malicious\nbehavior. The dataset has been collected in cooperation between Avast Software\nand Czech Technical University - AI Center (AIC).",
"title": "Avast-CTU Public CAPE Dataset",
"url": "http://arxiv.org/abs/2209.03188v1"
} | null | null | new_dataset | admin | null | false | null | 96829754-3cc8-440a-b232-1ca48c2a3c34 | null | Validated | {
"text_length": 1172
} | 0new_dataset
|
TITLE: RDD2022: A multi-national image dataset for automatic Road Damage Detection
ABSTRACT: The data article describes the Road Damage Dataset, RDD2022, which comprises
47,420 road images from six countries, Japan, India, the Czech Republic,
Norway, the United States, and China. The images have been annotated with more
than 55,000 instances of road damage. Four types of road damage, namely
longitudinal cracks, transverse cracks, alligator cracks, and potholes, are
captured in the dataset. The annotated dataset is envisioned for developing
deep learning-based methods to detect and classify road damage automatically.
The dataset has been released as a part of the Crowd sensing-based Road Damage
Detection Challenge (CRDDC2022). The challenge CRDDC2022 invites researchers
from across the globe to propose solutions for automatic road damage detection
in multiple countries. The municipalities and road agencies may utilize the
RDD2022 dataset, and the models trained using RDD2022 for low-cost automatic
monitoring of road conditions. Further, computer vision and machine learning
researchers may use the dataset to benchmark the performance of different
algorithms for other image-based applications of the same type (classification,
object detection, etc.). | {
"abstract": "The data article describes the Road Damage Dataset, RDD2022, which comprises\n47,420 road images from six countries, Japan, India, the Czech Republic,\nNorway, the United States, and China. The images have been annotated with more\nthan 55,000 instances of road damage. Four types of road damage, namely\nlongitudinal cracks, transverse cracks, alligator cracks, and potholes, are\ncaptured in the dataset. The annotated dataset is envisioned for developing\ndeep learning-based methods to detect and classify road damage automatically.\nThe dataset has been released as a part of the Crowd sensing-based Road Damage\nDetection Challenge (CRDDC2022). The challenge CRDDC2022 invites researchers\nfrom across the globe to propose solutions for automatic road damage detection\nin multiple countries. The municipalities and road agencies may utilize the\nRDD2022 dataset, and the models trained using RDD2022 for low-cost automatic\nmonitoring of road conditions. Further, computer vision and machine learning\nresearchers may use the dataset to benchmark the performance of different\nalgorithms for other image-based applications of the same type (classification,\nobject detection, etc.).",
"title": "RDD2022: A multi-national image dataset for automatic Road Damage Detection",
"url": "http://arxiv.org/abs/2209.08538v1"
} | null | null | new_dataset | admin | null | false | null | c0bcb794-f281-44ee-819a-6bb0253283a0 | null | Validated | {
"text_length": 1284
} | 0new_dataset
|
TITLE: AVOIDDS: Aircraft Vision-based Intruder Detection Dataset and Simulator
ABSTRACT: Designing robust machine learning systems remains an open problem, and there
is a need for benchmark problems that cover both environmental changes and
evaluation on a downstream task. In this work, we introduce AVOIDDS, a
realistic object detection benchmark for the vision-based aircraft
detect-and-avoid problem. We provide a labeled dataset consisting of 72,000
photorealistic images of intruder aircraft with various lighting conditions,
weather conditions, relative geometries, and geographic locations. We also
provide an interface that evaluates trained models on slices of this dataset to
identify changes in performance with respect to changing environmental
conditions. Finally, we implement a fully-integrated, closed-loop simulator of
the vision-based detect-and-avoid problem to evaluate trained models with
respect to the downstream collision avoidance task. This benchmark will enable
further research in the design of robust machine learning systems for use in
safety-critical applications. The AVOIDDS dataset and code are publicly
available at
$\href{https://purl.stanford.edu/hj293cv5980}{purl.stanford.edu/hj293cv5980}$
and
$\href{https://github.com/sisl/VisionBasedAircraftDAA}{github.com/sisl/VisionBasedAircraftDAA}$,
respectively. | {
"abstract": "Designing robust machine learning systems remains an open problem, and there\nis a need for benchmark problems that cover both environmental changes and\nevaluation on a downstream task. In this work, we introduce AVOIDDS, a\nrealistic object detection benchmark for the vision-based aircraft\ndetect-and-avoid problem. We provide a labeled dataset consisting of 72,000\nphotorealistic images of intruder aircraft with various lighting conditions,\nweather conditions, relative geometries, and geographic locations. We also\nprovide an interface that evaluates trained models on slices of this dataset to\nidentify changes in performance with respect to changing environmental\nconditions. Finally, we implement a fully-integrated, closed-loop simulator of\nthe vision-based detect-and-avoid problem to evaluate trained models with\nrespect to the downstream collision avoidance task. This benchmark will enable\nfurther research in the design of robust machine learning systems for use in\nsafety-critical applications. The AVOIDDS dataset and code are publicly\navailable at\n$\\href{https://purl.stanford.edu/hj293cv5980}{purl.stanford.edu/hj293cv5980}$\nand\n$\\href{https://github.com/sisl/VisionBasedAircraftDAA}{github.com/sisl/VisionBasedAircraftDAA}$,\nrespectively.",
"title": "AVOIDDS: Aircraft Vision-based Intruder Detection Dataset and Simulator",
"url": "http://arxiv.org/abs/2306.11203v1"
} | null | null | new_dataset | admin | null | false | null | 1d5acbd8-5051-49fe-b87f-c90562bbe35f | null | Validated | {
"text_length": 1361
} | 0new_dataset
|
TITLE: COMPASS: A Formal Framework and Aggregate Dataset for Generalized Surgical Procedure Modeling
ABSTRACT: Purpose: We propose a formal framework for the modeling and segmentation of
minimally-invasive surgical tasks using a unified set of motion primitives
(MPs) to enable more objective labeling and the aggregation of different
datasets.
Methods: We model dry-lab surgical tasks as finite state machines,
representing how the execution of MPs as the basic surgical actions results in
the change of surgical context, which characterizes the physical interactions
among tools and objects in the surgical environment. We develop methods for
labeling surgical context based on video data and for automatic translation of
context to MP labels. We then use our framework to create the COntext and
Motion Primitive Aggregate Surgical Set (COMPASS), including six dry-lab
surgical tasks from three publicly-available datasets (JIGSAWS, DESK, and
ROSMA), with kinematic and video data and context and MP labels.
Results: Our context labeling method achieves near-perfect agreement between
consensus labels from crowd-sourcing and expert surgeons. Segmentation of tasks
to MPs results in the creation of the COMPASS dataset that nearly triples the
amount of data for modeling and analysis and enables the generation of separate
transcripts for the left and right tools.
Conclusion: The proposed framework results in high quality labeling of
surgical data based on context and fine-grained MPs. Modeling surgical tasks
with MPs enables the aggregation of different datasets and the separate
analysis of left and right hands for bimanual coordination assessment. Our
formal framework and aggregate dataset can support the development of
explainable and multi-granularity models for improved surgical process
analysis, skill assessment, error detection, and autonomy. | {
"abstract": "Purpose: We propose a formal framework for the modeling and segmentation of\nminimally-invasive surgical tasks using a unified set of motion primitives\n(MPs) to enable more objective labeling and the aggregation of different\ndatasets.\n Methods: We model dry-lab surgical tasks as finite state machines,\nrepresenting how the execution of MPs as the basic surgical actions results in\nthe change of surgical context, which characterizes the physical interactions\namong tools and objects in the surgical environment. We develop methods for\nlabeling surgical context based on video data and for automatic translation of\ncontext to MP labels. We then use our framework to create the COntext and\nMotion Primitive Aggregate Surgical Set (COMPASS), including six dry-lab\nsurgical tasks from three publicly-available datasets (JIGSAWS, DESK, and\nROSMA), with kinematic and video data and context and MP labels.\n Results: Our context labeling method achieves near-perfect agreement between\nconsensus labels from crowd-sourcing and expert surgeons. Segmentation of tasks\nto MPs results in the creation of the COMPASS dataset that nearly triples the\namount of data for modeling and analysis and enables the generation of separate\ntranscripts for the left and right tools.\n Conclusion: The proposed framework results in high quality labeling of\nsurgical data based on context and fine-grained MPs. Modeling surgical tasks\nwith MPs enables the aggregation of different datasets and the separate\nanalysis of left and right hands for bimanual coordination assessment. Our\nformal framework and aggregate dataset can support the development of\nexplainable and multi-granularity models for improved surgical process\nanalysis, skill assessment, error detection, and autonomy.",
"title": "COMPASS: A Formal Framework and Aggregate Dataset for Generalized Surgical Procedure Modeling",
"url": "http://arxiv.org/abs/2209.06424v5"
} | null | null | new_dataset | admin | null | false | null | 7ddec2dd-89f4-475d-9328-c02c883bac86 | null | Validated | {
"text_length": 1884
} | 0new_dataset
|
TITLE: DynaBench: A benchmark dataset for learning dynamical systems from low-resolution data
ABSTRACT: Previous work on learning physical systems from data has focused on
high-resolution grid-structured measurements. However, real-world knowledge of
such systems (e.g. weather data) relies on sparsely scattered measuring
stations. In this paper, we introduce a novel simulated benchmark dataset,
DynaBench, for learning dynamical systems directly from sparsely scattered data
without prior knowledge of the equations. The dataset focuses on predicting the
evolution of a dynamical system from low-resolution, unstructured measurements.
We simulate six different partial differential equations covering a variety of
physical systems commonly used in the literature and evaluate several machine
learning models, including traditional graph neural networks and point cloud
processing models, with the task of predicting the evolution of the system. The
proposed benchmark dataset is expected to advance the state of art as an
out-of-the-box easy-to-use tool for evaluating models in a setting where only
unstructured low-resolution observations are available. The benchmark is
available at https://anonymous.4open.science/r/code-2022-dynabench/. | {
"abstract": "Previous work on learning physical systems from data has focused on\nhigh-resolution grid-structured measurements. However, real-world knowledge of\nsuch systems (e.g. weather data) relies on sparsely scattered measuring\nstations. In this paper, we introduce a novel simulated benchmark dataset,\nDynaBench, for learning dynamical systems directly from sparsely scattered data\nwithout prior knowledge of the equations. The dataset focuses on predicting the\nevolution of a dynamical system from low-resolution, unstructured measurements.\nWe simulate six different partial differential equations covering a variety of\nphysical systems commonly used in the literature and evaluate several machine\nlearning models, including traditional graph neural networks and point cloud\nprocessing models, with the task of predicting the evolution of the system. The\nproposed benchmark dataset is expected to advance the state of art as an\nout-of-the-box easy-to-use tool for evaluating models in a setting where only\nunstructured low-resolution observations are available. The benchmark is\navailable at https://anonymous.4open.science/r/code-2022-dynabench/.",
"title": "DynaBench: A benchmark dataset for learning dynamical systems from low-resolution data",
"url": "http://arxiv.org/abs/2306.05805v2"
} | null | null | new_dataset | admin | null | false | null | ea320f3e-d049-455e-adcc-bfbdd1fa8a2b | null | Validated | {
"text_length": 1261
} | 0new_dataset
|
TITLE: Massive MIMO Channel Prediction Via Meta-Learning and Deep Denoising: Is a Small Dataset Enough?
ABSTRACT: Accurate channel knowledge is critical in massive multiple-input
multiple-output (MIMO), which motivates the use of channel prediction. Machine
learning techniques for channel prediction hold much promise, but current
schemes are limited in their ability to adapt to changes in the environment
because they require large training overheads. To accurately predict wireless
channels for new environments with reduced training overhead, we propose a fast
adaptive channel prediction technique based on a meta-learning algorithm for
massive MIMO communications. We exploit the model-agnostic meta-learning (MAML)
algorithm to achieve quick adaptation with a small amount of labeled data.
Also, to improve the prediction accuracy, we adopt the denoising process for
the training data by using deep image prior (DIP). Numerical results show that
the proposed MAML-based channel predictor can improve the prediction accuracy
with only a few fine-tuning samples. The DIP-based denoising process gives an
additional gain in channel prediction, especially in low signal-to-noise ratio
regimes. | {
"abstract": "Accurate channel knowledge is critical in massive multiple-input\nmultiple-output (MIMO), which motivates the use of channel prediction. Machine\nlearning techniques for channel prediction hold much promise, but current\nschemes are limited in their ability to adapt to changes in the environment\nbecause they require large training overheads. To accurately predict wireless\nchannels for new environments with reduced training overhead, we propose a fast\nadaptive channel prediction technique based on a meta-learning algorithm for\nmassive MIMO communications. We exploit the model-agnostic meta-learning (MAML)\nalgorithm to achieve quick adaptation with a small amount of labeled data.\nAlso, to improve the prediction accuracy, we adopt the denoising process for\nthe training data by using deep image prior (DIP). Numerical results show that\nthe proposed MAML-based channel predictor can improve the prediction accuracy\nwith only a few fine-tuning samples. The DIP-based denoising process gives an\nadditional gain in channel prediction, especially in low signal-to-noise ratio\nregimes.",
"title": "Massive MIMO Channel Prediction Via Meta-Learning and Deep Denoising: Is a Small Dataset Enough?",
"url": "http://arxiv.org/abs/2210.08770v1"
} | null | null | no_new_dataset | admin | null | false | null | 60b9cb2d-7bc6-46ad-bc59-6a260bc9a069 | null | Validated | {
"text_length": 1214
} | 1no_new_dataset
|
TITLE: TTSWING: a Dataset for Table Tennis Swing Analysis
ABSTRACT: We introduce TTSWING, a novel dataset designed for table tennis swing
analysis. This dataset comprises comprehensive swing information obtained
through 9-axis sensors integrated into custom-made racket grips, accompanied by
anonymized demographic data of the players. We detail the data collection and
annotation procedures. Furthermore, we conduct pilot studies utilizing diverse
machine learning models for swing analysis. TTSWING holds tremendous potential
to facilitate innovative research in table tennis analysis and is a valuable
resource for the scientific community. We release the dataset and experimental
codes at https://github.com/DEPhantom/TTSWING. | {
"abstract": "We introduce TTSWING, a novel dataset designed for table tennis swing\nanalysis. This dataset comprises comprehensive swing information obtained\nthrough 9-axis sensors integrated into custom-made racket grips, accompanied by\nanonymized demographic data of the players. We detail the data collection and\nannotation procedures. Furthermore, we conduct pilot studies utilizing diverse\nmachine learning models for swing analysis. TTSWING holds tremendous potential\nto facilitate innovative research in table tennis analysis and is a valuable\nresource for the scientific community. We release the dataset and experimental\ncodes at https://github.com/DEPhantom/TTSWING.",
"title": "TTSWING: a Dataset for Table Tennis Swing Analysis",
"url": "http://arxiv.org/abs/2306.17550v1"
} | null | null | new_dataset | admin | null | false | null | 1279b7a2-2a6a-437b-ad7e-bc5a87c03f53 | null | Validated | {
"text_length": 747
} | 0new_dataset
|
TITLE: Evaluating and Crafting Datasets Effective for Deep Learning With Data Maps
ABSTRACT: Rapid development in deep learning model construction has prompted an
increased need for appropriate training data. The popularity of large datasets
- sometimes known as "big data" - has diverted attention from assessing their
quality. Training on large datasets often requires excessive system resources
and an infeasible amount of time. Furthermore, the supervised machine learning
process has yet to be fully automated: for supervised learning, large datasets
require more time for manually labeling samples. We propose a method of
curating smaller datasets with comparable out-of-distribution model accuracy
after an initial training session using an appropriate distribution of samples
classified by how difficult it is for a model to learn from them. | {
"abstract": "Rapid development in deep learning model construction has prompted an\nincreased need for appropriate training data. The popularity of large datasets\n- sometimes known as \"big data\" - has diverted attention from assessing their\nquality. Training on large datasets often requires excessive system resources\nand an infeasible amount of time. Furthermore, the supervised machine learning\nprocess has yet to be fully automated: for supervised learning, large datasets\nrequire more time for manually labeling samples. We propose a method of\ncurating smaller datasets with comparable out-of-distribution model accuracy\nafter an initial training session using an appropriate distribution of samples\nclassified by how difficult it is for a model to learn from them.",
"title": "Evaluating and Crafting Datasets Effective for Deep Learning With Data Maps",
"url": "http://arxiv.org/abs/2208.10033v2"
} | null | null | no_new_dataset | admin | null | false | null | 35163b28-914d-472d-b705-15c31d09c376 | null | Validated | {
"text_length": 866
} | 1no_new_dataset
|
TITLE: Dataset Inference: Ownership Resolution in Machine Learning
ABSTRACT: With increasingly more data and computation involved in their training,
machine learning models constitute valuable intellectual property. This has
spurred interest in model stealing, which is made more practical by advances in
learning with partial, little, or no supervision. Existing defenses focus on
inserting unique watermarks in a model's decision surface, but this is
insufficient: the watermarks are not sampled from the training distribution and
thus are not always preserved during model stealing. In this paper, we make the
key observation that knowledge contained in the stolen model's training set is
what is common to all stolen copies. The adversary's goal, irrespective of the
attack employed, is always to extract this knowledge or its by-products. This
gives the original model's owner a strong advantage over the adversary: model
owners have access to the original training data. We thus introduce $dataset$
$inference$, the process of identifying whether a suspected model copy has
private knowledge from the original model's dataset, as a defense against model
stealing. We develop an approach for dataset inference that combines
statistical testing with the ability to estimate the distance of multiple data
points to the decision boundary. Our experiments on CIFAR10, SVHN, CIFAR100 and
ImageNet show that model owners can claim with confidence greater than 99% that
their model (or dataset as a matter of fact) was stolen, despite only exposing
50 of the stolen model's training points. Dataset inference defends against
state-of-the-art attacks even when the adversary is adaptive. Unlike prior
work, it does not require retraining or overfitting the defended model. | {
"abstract": "With increasingly more data and computation involved in their training,\nmachine learning models constitute valuable intellectual property. This has\nspurred interest in model stealing, which is made more practical by advances in\nlearning with partial, little, or no supervision. Existing defenses focus on\ninserting unique watermarks in a model's decision surface, but this is\ninsufficient: the watermarks are not sampled from the training distribution and\nthus are not always preserved during model stealing. In this paper, we make the\nkey observation that knowledge contained in the stolen model's training set is\nwhat is common to all stolen copies. The adversary's goal, irrespective of the\nattack employed, is always to extract this knowledge or its by-products. This\ngives the original model's owner a strong advantage over the adversary: model\nowners have access to the original training data. We thus introduce $dataset$\n$inference$, the process of identifying whether a suspected model copy has\nprivate knowledge from the original model's dataset, as a defense against model\nstealing. We develop an approach for dataset inference that combines\nstatistical testing with the ability to estimate the distance of multiple data\npoints to the decision boundary. Our experiments on CIFAR10, SVHN, CIFAR100 and\nImageNet show that model owners can claim with confidence greater than 99% that\ntheir model (or dataset as a matter of fact) was stolen, despite only exposing\n50 of the stolen model's training points. Dataset inference defends against\nstate-of-the-art attacks even when the adversary is adaptive. Unlike prior\nwork, it does not require retraining or overfitting the defended model.",
"title": "Dataset Inference: Ownership Resolution in Machine Learning",
"url": "http://arxiv.org/abs/2104.10706v1"
} | null | null | no_new_dataset | admin | null | false | null | 34e24349-1528-48dd-a7cc-b25e36470f4d | null | Validated | {
"text_length": 1786
} | 1no_new_dataset
|
TITLE: A Benchmark dataset for predictive maintenance
ABSTRACT: The paper describes the MetroPT data set, an outcome of a eXplainable
Predictive Maintenance (XPM) project with an urban metro public transportation
service in Porto, Portugal. The data was collected in 2022 that aimed to
evaluate machine learning methods for online anomaly detection and failure
prediction. By capturing several analogic sensor signals (pressure,
temperature, current consumption), digital signals (control signals, discrete
signals), and GPS information (latitude, longitude, and speed), we provide a
dataset that can be easily used to evaluate online machine learning methods.
This dataset contains some interesting characteristics and can be a good
benchmark for predictive maintenance models. | {
"abstract": "The paper describes the MetroPT data set, an outcome of a eXplainable\nPredictive Maintenance (XPM) project with an urban metro public transportation\nservice in Porto, Portugal. The data was collected in 2022 that aimed to\nevaluate machine learning methods for online anomaly detection and failure\nprediction. By capturing several analogic sensor signals (pressure,\ntemperature, current consumption), digital signals (control signals, discrete\nsignals), and GPS information (latitude, longitude, and speed), we provide a\ndataset that can be easily used to evaluate online machine learning methods.\nThis dataset contains some interesting characteristics and can be a good\nbenchmark for predictive maintenance models.",
"title": "A Benchmark dataset for predictive maintenance",
"url": "http://arxiv.org/abs/2207.05466v3"
} | null | null | new_dataset | admin | null | false | null | 35d2af1e-7208-481b-b3eb-491c6f840920 | null | Validated | {
"text_length": 795
} | 0new_dataset
|
TITLE: Data Feedback Loops: Model-driven Amplification of Dataset Biases
ABSTRACT: Datasets scraped from the internet have been critical to the successes of
large-scale machine learning. Yet, this very success puts the utility of future
internet-derived datasets at potential risk, as model outputs begin to replace
human annotations as a source of supervision.
In this work, we first formalize a system where interactions with one model
are recorded as history and scraped as training data in the future. We then
analyze its stability over time by tracking changes to a test-time bias
statistic (e.g. gender bias of model predictions). We find that the degree of
bias amplification is closely linked to whether the model's outputs behave like
samples from the training distribution, a behavior which we characterize and
define as consistent calibration. Experiments in three conditional prediction
scenarios - image classification, visual role-labeling, and language generation
- demonstrate that models that exhibit a sampling-like behavior are more
calibrated and thus more stable. Based on this insight, we propose an
intervention to help calibrate and stabilize unstable feedback systems.
Code is available at https://github.com/rtaori/data_feedback. | {
"abstract": "Datasets scraped from the internet have been critical to the successes of\nlarge-scale machine learning. Yet, this very success puts the utility of future\ninternet-derived datasets at potential risk, as model outputs begin to replace\nhuman annotations as a source of supervision.\n In this work, we first formalize a system where interactions with one model\nare recorded as history and scraped as training data in the future. We then\nanalyze its stability over time by tracking changes to a test-time bias\nstatistic (e.g. gender bias of model predictions). We find that the degree of\nbias amplification is closely linked to whether the model's outputs behave like\nsamples from the training distribution, a behavior which we characterize and\ndefine as consistent calibration. Experiments in three conditional prediction\nscenarios - image classification, visual role-labeling, and language generation\n- demonstrate that models that exhibit a sampling-like behavior are more\ncalibrated and thus more stable. Based on this insight, we propose an\nintervention to help calibrate and stabilize unstable feedback systems.\n Code is available at https://github.com/rtaori/data_feedback.",
"title": "Data Feedback Loops: Model-driven Amplification of Dataset Biases",
"url": "http://arxiv.org/abs/2209.03942v1"
} | null | null | no_new_dataset | admin | null | false | null | e56879a8-ab90-4429-ad5e-20889e984deb | null | Validated | {
"text_length": 1276
} | 1no_new_dataset
|
TITLE: ClimART: A Benchmark Dataset for Emulating Atmospheric Radiative Transfer in Weather and Climate Models
ABSTRACT: Numerical simulations of Earth's weather and climate require substantial
amounts of computation. This has led to a growing interest in replacing
subroutines that explicitly compute physical processes with approximate machine
learning (ML) methods that are fast at inference time. Within weather and
climate models, atmospheric radiative transfer (RT) calculations are especially
expensive. This has made them a popular target for neural network-based
emulators. However, prior work is hard to compare due to the lack of a
comprehensive dataset and standardized best practices for ML benchmarking. To
fill this gap, we build a large dataset, ClimART, with more than \emph{10
million samples from present, pre-industrial, and future climate conditions},
based on the Canadian Earth System Model. ClimART poses several methodological
challenges for the ML community, such as multiple out-of-distribution test
sets, underlying domain physics, and a trade-off between accuracy and inference
speed. We also present several novel baselines that indicate shortcomings of
datasets and network architectures used in prior work. Download instructions,
baselines, and code are available at: https://github.com/RolnickLab/climart | {
"abstract": "Numerical simulations of Earth's weather and climate require substantial\namounts of computation. This has led to a growing interest in replacing\nsubroutines that explicitly compute physical processes with approximate machine\nlearning (ML) methods that are fast at inference time. Within weather and\nclimate models, atmospheric radiative transfer (RT) calculations are especially\nexpensive. This has made them a popular target for neural network-based\nemulators. However, prior work is hard to compare due to the lack of a\ncomprehensive dataset and standardized best practices for ML benchmarking. To\nfill this gap, we build a large dataset, ClimART, with more than \\emph{10\nmillion samples from present, pre-industrial, and future climate conditions},\nbased on the Canadian Earth System Model. ClimART poses several methodological\nchallenges for the ML community, such as multiple out-of-distribution test\nsets, underlying domain physics, and a trade-off between accuracy and inference\nspeed. We also present several novel baselines that indicate shortcomings of\ndatasets and network architectures used in prior work. Download instructions,\nbaselines, and code are available at: https://github.com/RolnickLab/climart",
"title": "ClimART: A Benchmark Dataset for Emulating Atmospheric Radiative Transfer in Weather and Climate Models",
"url": "http://arxiv.org/abs/2111.14671v1"
} | null | null | new_dataset | admin | null | false | null | 279620ec-4c31-4ecb-9f13-d91a3b39ebe6 | null | Validated | {
"text_length": 1354
} | 0new_dataset
|
TITLE: Critical Evaluation of LOCO dataset with Machine Learning
ABSTRACT: Purpose: Object detection is rapidly evolving through machine learning
technology in automation systems. Well prepared data is necessary to train the
algorithms. Accordingly, the objective of this paper is to describe a
re-evaluation of the so-called Logistics Objects in Context (LOCO) dataset,
which is the first dataset for object detection in the field of intralogistics.
Methodology: We use an experimental research approach with three steps to
evaluate the LOCO dataset. Firstly, the images on GitHub were analyzed to
understand the dataset better. Secondly, Google Drive Cloud was used for
training purposes to revisit the algorithmic implementation and training.
Lastly, the LOCO dataset was examined, if it is possible to achieve the same
training results in comparison to the original publications.
Findings: The mean average precision, a common benchmark in object detection,
achieved in our study was 64.54%, and shows a significant increase from the
initial study of the LOCO authors, achieving 41%. However, improvement
potential is seen specifically within object types of forklifts and pallet
truck.
Originality: This paper presents the first critical replication study of the
LOCO dataset for object detection in intralogistics. It shows that the training
with better hyperparameters based on LOCO can even achieve a higher accuracy
than presented in the original publication. However, there is also further room
for improving the LOCO dataset. | {
"abstract": "Purpose: Object detection is rapidly evolving through machine learning\ntechnology in automation systems. Well prepared data is necessary to train the\nalgorithms. Accordingly, the objective of this paper is to describe a\nre-evaluation of the so-called Logistics Objects in Context (LOCO) dataset,\nwhich is the first dataset for object detection in the field of intralogistics.\n Methodology: We use an experimental research approach with three steps to\nevaluate the LOCO dataset. Firstly, the images on GitHub were analyzed to\nunderstand the dataset better. Secondly, Google Drive Cloud was used for\ntraining purposes to revisit the algorithmic implementation and training.\nLastly, the LOCO dataset was examined, if it is possible to achieve the same\ntraining results in comparison to the original publications.\n Findings: The mean average precision, a common benchmark in object detection,\nachieved in our study was 64.54%, and shows a significant increase from the\ninitial study of the LOCO authors, achieving 41%. However, improvement\npotential is seen specifically within object types of forklifts and pallet\ntruck.\n Originality: This paper presents the first critical replication study of the\nLOCO dataset for object detection in intralogistics. It shows that the training\nwith better hyperparameters based on LOCO can even achieve a higher accuracy\nthan presented in the original publication. However, there is also further room\nfor improving the LOCO dataset.",
"title": "Critical Evaluation of LOCO dataset with Machine Learning",
"url": "http://arxiv.org/abs/2209.13499v1"
} | null | null | no_new_dataset | admin | null | false | null | 7892fa27-e42b-4d96-936e-9fd751f74bbe | null | Validated | {
"text_length": 1559
} | 1no_new_dataset
|
TITLE: Architectural Optimization and Feature Learning for High-Dimensional Time Series Datasets
ABSTRACT: As our ability to sense increases, we are experiencing a transition from
data-poor problems, in which the central issue is a lack of relevant data, to
data-rich problems, in which the central issue is to identify a few relevant
features in a sea of observations. Motivated by applications in
gravitational-wave astrophysics, we study the problem of predicting the
presence of transient noise artifacts in a gravitational wave detector from a
rich collection of measurements from the detector and its environment. We argue
that feature learning--in which relevant features are optimized from data--is
critical to achieving high accuracy. We introduce models that reduce the error
rate by over 60% compared to the previous state of the art, which used fixed,
hand-crafted features. Feature learning is useful not only because it improves
performance on prediction tasks; the results provide valuable information about
patterns associated with phenomena of interest that would otherwise be
undiscoverable. In our application, features found to be associated with
transient noise provide diagnostic information about its origin and suggest
mitigation strategies. Learning in high-dimensional settings is challenging.
Through experiments with a variety of architectures, we identify two key
factors in successful models: sparsity, for selecting relevant variables within
the high-dimensional observations; and depth, which confers flexibility for
handling complex interactions and robustness with respect to temporal
variations. We illustrate their significance through systematic experiments on
real detector data. Our results provide experimental corroboration of common
assumptions in the machine-learning community and have direct applicability to
improving our ability to sense gravitational waves, as well as to many other
problem settings with similarly high-dimensional, noisy, or partly irrelevant
data. | {
"abstract": "As our ability to sense increases, we are experiencing a transition from\ndata-poor problems, in which the central issue is a lack of relevant data, to\ndata-rich problems, in which the central issue is to identify a few relevant\nfeatures in a sea of observations. Motivated by applications in\ngravitational-wave astrophysics, we study the problem of predicting the\npresence of transient noise artifacts in a gravitational wave detector from a\nrich collection of measurements from the detector and its environment. We argue\nthat feature learning--in which relevant features are optimized from data--is\ncritical to achieving high accuracy. We introduce models that reduce the error\nrate by over 60% compared to the previous state of the art, which used fixed,\nhand-crafted features. Feature learning is useful not only because it improves\nperformance on prediction tasks; the results provide valuable information about\npatterns associated with phenomena of interest that would otherwise be\nundiscoverable. In our application, features found to be associated with\ntransient noise provide diagnostic information about its origin and suggest\nmitigation strategies. Learning in high-dimensional settings is challenging.\nThrough experiments with a variety of architectures, we identify two key\nfactors in successful models: sparsity, for selecting relevant variables within\nthe high-dimensional observations; and depth, which confers flexibility for\nhandling complex interactions and robustness with respect to temporal\nvariations. We illustrate their significance through systematic experiments on\nreal detector data. Our results provide experimental corroboration of common\nassumptions in the machine-learning community and have direct applicability to\nimproving our ability to sense gravitational waves, as well as to many other\nproblem settings with similarly high-dimensional, noisy, or partly irrelevant\ndata.",
"title": "Architectural Optimization and Feature Learning for High-Dimensional Time Series Datasets",
"url": "http://arxiv.org/abs/2202.13486v2"
} | null | null | no_new_dataset | admin | null | false | null | 354aa25b-0f94-4d9e-b412-a830e2809237 | null | Validated | {
"text_length": 2031
} | 1no_new_dataset
|
TITLE: Potrika: Raw and Balanced Newspaper Datasets in the Bangla Language with Eight Topics and Five Attributes
ABSTRACT: Knowledge is central to human and scientific developments. Natural Language
Processing (NLP) allows automated analysis and creation of knowledge. Data is a
crucial NLP and machine learning ingredient. The scarcity of open datasets is a
well-known problem in machine and deep learning research. This is very much the
case for textual NLP datasets in English and other major world languages. For
the Bangla language, the situation is even more challenging and the number of
large datasets for NLP research is practically nil. We hereby present Potrika,
a large single-label Bangla news article textual dataset curated for NLP
research from six popular online news portals in Bangladesh (Jugantor,
Jaijaidin, Ittefaq, Kaler Kontho, Inqilab, and Somoyer Alo) for the period
2014-2020. The articles are classified into eight distinct categories
(National, Sports, International, Entertainment, Economy, Education, Politics,
and Science \& Technology) providing five attributes (News Article, Category,
Headline, Publication Date, and Newspaper Source). The raw dataset contains
185.51 million words and 12.57 million sentences contained in 664,880 news
articles. Moreover, using NLP augmentation techniques, we create from the raw
(unbalanced) dataset another (balanced) dataset comprising 320,000 news
articles with 40,000 articles in each of the eight news categories. Potrika
contains both the datasets (raw and balanced) to suit a wide range of NLP
research. By far, to the best of our knowledge, Potrika is the largest and the
most extensive dataset for news classification. | {
"abstract": "Knowledge is central to human and scientific developments. Natural Language\nProcessing (NLP) allows automated analysis and creation of knowledge. Data is a\ncrucial NLP and machine learning ingredient. The scarcity of open datasets is a\nwell-known problem in machine and deep learning research. This is very much the\ncase for textual NLP datasets in English and other major world languages. For\nthe Bangla language, the situation is even more challenging and the number of\nlarge datasets for NLP research is practically nil. We hereby present Potrika,\na large single-label Bangla news article textual dataset curated for NLP\nresearch from six popular online news portals in Bangladesh (Jugantor,\nJaijaidin, Ittefaq, Kaler Kontho, Inqilab, and Somoyer Alo) for the period\n2014-2020. The articles are classified into eight distinct categories\n(National, Sports, International, Entertainment, Economy, Education, Politics,\nand Science \\& Technology) providing five attributes (News Article, Category,\nHeadline, Publication Date, and Newspaper Source). The raw dataset contains\n185.51 million words and 12.57 million sentences contained in 664,880 news\narticles. Moreover, using NLP augmentation techniques, we create from the raw\n(unbalanced) dataset another (balanced) dataset comprising 320,000 news\narticles with 40,000 articles in each of the eight news categories. Potrika\ncontains both the datasets (raw and balanced) to suit a wide range of NLP\nresearch. By far, to the best of our knowledge, Potrika is the largest and the\nmost extensive dataset for news classification.",
"title": "Potrika: Raw and Balanced Newspaper Datasets in the Bangla Language with Eight Topics and Five Attributes",
"url": "http://arxiv.org/abs/2210.09389v1"
} | null | null | new_dataset | admin | null | false | null | 7fe08fc6-104f-4522-b868-4114f14aa33c | null | Validated | {
"text_length": 1714
} | 0new_dataset
|
TITLE: The Open Catalyst 2022 (OC22) Dataset and Challenges for Oxide Electrocatalysts
ABSTRACT: The development of machine learning models for electrocatalysts requires a
broad set of training data to enable their use across a wide variety of
materials. One class of materials that currently lacks sufficient training data
is oxides, which are critical for the development of OER catalysts. To address
this, we developed the OC22 dataset, consisting of 62,331 DFT relaxations
(~9,854,504 single point calculations) across a range of oxide materials,
coverages, and adsorbates. We define generalized total energy tasks that enable
property prediction beyond adsorption energies; we test baseline performance of
several graph neural networks; and we provide pre-defined dataset splits to
establish clear benchmarks for future efforts. In the most general task,
GemNet-OC sees a ~36% improvement in energy predictions when combining the
chemically dissimilar OC20 and OC22 datasets via fine-tuning. Similarly, we
achieved a ~19% improvement in total energy predictions on OC20 and a ~9%
improvement in force predictions in OC22 when using joint training. We
demonstrate the practical utility of a top performing model by capturing
literature adsorption energies and important OER scaling relationships. We
expect OC22 to provide an important benchmark for models seeking to incorporate
intricate long-range electrostatic and magnetic interactions in oxide surfaces.
Dataset and baseline models are open sourced, and a public leaderboard is
available to encourage continued community developments on the total energy
tasks and data. | {
"abstract": "The development of machine learning models for electrocatalysts requires a\nbroad set of training data to enable their use across a wide variety of\nmaterials. One class of materials that currently lacks sufficient training data\nis oxides, which are critical for the development of OER catalysts. To address\nthis, we developed the OC22 dataset, consisting of 62,331 DFT relaxations\n(~9,854,504 single point calculations) across a range of oxide materials,\ncoverages, and adsorbates. We define generalized total energy tasks that enable\nproperty prediction beyond adsorption energies; we test baseline performance of\nseveral graph neural networks; and we provide pre-defined dataset splits to\nestablish clear benchmarks for future efforts. In the most general task,\nGemNet-OC sees a ~36% improvement in energy predictions when combining the\nchemically dissimilar OC20 and OC22 datasets via fine-tuning. Similarly, we\nachieved a ~19% improvement in total energy predictions on OC20 and a ~9%\nimprovement in force predictions in OC22 when using joint training. We\ndemonstrate the practical utility of a top performing model by capturing\nliterature adsorption energies and important OER scaling relationships. We\nexpect OC22 to provide an important benchmark for models seeking to incorporate\nintricate long-range electrostatic and magnetic interactions in oxide surfaces.\nDataset and baseline models are open sourced, and a public leaderboard is\navailable to encourage continued community developments on the total energy\ntasks and data.",
"title": "The Open Catalyst 2022 (OC22) Dataset and Challenges for Oxide Electrocatalysts",
"url": "http://arxiv.org/abs/2206.08917v3"
} | null | null | new_dataset | admin | null | false | null | 165e43bb-eb0b-4bc5-a64e-a4cdf40f12c2 | null | Validated | {
"text_length": 1646
} | 0new_dataset
|
TITLE: BrazilDAM: A Benchmark dataset for Tailings Dam Detection
ABSTRACT: In this work we present BrazilDAM, a novel public dataset based on Sentinel-2
and Landsat-8 satellite images covering all tailings dams cataloged by the
Brazilian National Mining Agency (ANM). The dataset was built using
georeferenced images from 769 dams, recorded between 2016 and 2019. The time
series were processed in order to produce cloud free images. The dams contain
mining waste from different ore categories and have highly varying shapes,
areas and volumes, making BrazilDAM particularly interesting and challenging to
be used in machine learning benchmarks. The original catalog contains, besides
the dam coordinates, information about: the main ore, constructive method, risk
category, and associated potential damage. To evaluate BrazilDAM's predictive
potential we performed classification essays using state-of-the-art deep
Convolutional Neural Network (CNNs). In the experiments, we achieved an average
classification accuracy of 94.11% in tailing dam binary classification task. In
addition, others four setups of experiments were made using the complementary
information from the original catalog, exhaustively exploiting the capacity of
the proposed dataset. | {
"abstract": "In this work we present BrazilDAM, a novel public dataset based on Sentinel-2\nand Landsat-8 satellite images covering all tailings dams cataloged by the\nBrazilian National Mining Agency (ANM). The dataset was built using\ngeoreferenced images from 769 dams, recorded between 2016 and 2019. The time\nseries were processed in order to produce cloud free images. The dams contain\nmining waste from different ore categories and have highly varying shapes,\nareas and volumes, making BrazilDAM particularly interesting and challenging to\nbe used in machine learning benchmarks. The original catalog contains, besides\nthe dam coordinates, information about: the main ore, constructive method, risk\ncategory, and associated potential damage. To evaluate BrazilDAM's predictive\npotential we performed classification essays using state-of-the-art deep\nConvolutional Neural Network (CNNs). In the experiments, we achieved an average\nclassification accuracy of 94.11% in tailing dam binary classification task. In\naddition, others four setups of experiments were made using the complementary\ninformation from the original catalog, exhaustively exploiting the capacity of\nthe proposed dataset.",
"title": "BrazilDAM: A Benchmark dataset for Tailings Dam Detection",
"url": "http://arxiv.org/abs/2003.07948v2"
} | null | null | new_dataset | admin | null | false | null | df34d82e-ae49-42db-b2cc-daf65dee09d6 | null | Validated | {
"text_length": 1271
} | 0new_dataset
|
TITLE: Generative Adversarial Nets: Can we generate a new dataset based on only one training set?
ABSTRACT: A generative adversarial network (GAN) is a class of machine learning
frameworks designed by Goodfellow et al. in 2014. In the GAN framework, the
generative model is pitted against an adversary: a discriminative model that
learns to determine whether a sample is from the model distribution or the data
distribution. GAN generates new samples from the same distribution as the
training set. In this work, we aim to generate a new dataset that has a
different distribution from the training set. In addition, the Jensen-Shannon
divergence between the distributions of the generative and training datasets
can be controlled by some target $\delta \in [0, 1]$. Our work is motivated by
applications in generating new kinds of rice that have similar characteristics
as good rice. | {
"abstract": "A generative adversarial network (GAN) is a class of machine learning\nframeworks designed by Goodfellow et al. in 2014. In the GAN framework, the\ngenerative model is pitted against an adversary: a discriminative model that\nlearns to determine whether a sample is from the model distribution or the data\ndistribution. GAN generates new samples from the same distribution as the\ntraining set. In this work, we aim to generate a new dataset that has a\ndifferent distribution from the training set. In addition, the Jensen-Shannon\ndivergence between the distributions of the generative and training datasets\ncan be controlled by some target $\\delta \\in [0, 1]$. Our work is motivated by\napplications in generating new kinds of rice that have similar characteristics\nas good rice.",
"title": "Generative Adversarial Nets: Can we generate a new dataset based on only one training set?",
"url": "http://arxiv.org/abs/2210.06005v1"
} | null | null | no_new_dataset | admin | null | false | null | 52c20045-03c7-4b3d-a2db-b5f74019896c | null | Default | {
"text_length": 900
} | 1no_new_dataset
|
TITLE: Synthetic Dataset Generation for Privacy-Preserving Machine Learning
ABSTRACT: Machine Learning (ML) has achieved enormous success in solving a variety of
problems in computer vision, speech recognition, object detection, to name a
few. The principal reason for this success is the availability of huge datasets
for training deep neural networks (DNNs). However, datasets can not be publicly
released if they contain sensitive information such as medical or financial
records. In such cases, data privacy becomes a major concern. Encryption
methods offer a possible solution to this issue, however their deployment on ML
applications is non-trivial, as they seriously impact the classification
accuracy and result in substantial computational overhead.Alternatively,
obfuscation techniques can be used, but maintaining a good balance between
visual privacy and accuracy is challenging. In this work, we propose a method
to generate secure synthetic datasets from the original private datasets. In
our method, given a network with Batch Normalization (BN) layers pre-trained on
the original dataset, we first record the layer-wise BN statistics. Next, using
the BN statistics and the pre-trained model, we generate the synthetic dataset
by optimizing random noises such that the synthetic data match the layer-wise
statistical distribution of the original model. We evaluate our method on image
classification dataset (CIFAR10) and show that our synthetic data can be used
for training networks from scratch, producing reasonable classification
performance. | {
"abstract": "Machine Learning (ML) has achieved enormous success in solving a variety of\nproblems in computer vision, speech recognition, object detection, to name a\nfew. The principal reason for this success is the availability of huge datasets\nfor training deep neural networks (DNNs). However, datasets can not be publicly\nreleased if they contain sensitive information such as medical or financial\nrecords. In such cases, data privacy becomes a major concern. Encryption\nmethods offer a possible solution to this issue, however their deployment on ML\napplications is non-trivial, as they seriously impact the classification\naccuracy and result in substantial computational overhead.Alternatively,\nobfuscation techniques can be used, but maintaining a good balance between\nvisual privacy and accuracy is challenging. In this work, we propose a method\nto generate secure synthetic datasets from the original private datasets. In\nour method, given a network with Batch Normalization (BN) layers pre-trained on\nthe original dataset, we first record the layer-wise BN statistics. Next, using\nthe BN statistics and the pre-trained model, we generate the synthetic dataset\nby optimizing random noises such that the synthetic data match the layer-wise\nstatistical distribution of the original model. We evaluate our method on image\nclassification dataset (CIFAR10) and show that our synthetic data can be used\nfor training networks from scratch, producing reasonable classification\nperformance.",
"title": "Synthetic Dataset Generation for Privacy-Preserving Machine Learning",
"url": "http://arxiv.org/abs/2210.03205v5"
} | null | null | no_new_dataset | admin | null | false | null | 7aea2fea-96db-4732-8496-7f4da428882c | null | Validated | {
"text_length": 1580
} | 1no_new_dataset
|
TITLE: An Instance Selection Algorithm for Big Data in High imbalanced datasets based on LSH
ABSTRACT: Training of Machine Learning (ML) models in real contexts often deals with
big data sets and high-class imbalance samples where the class of interest is
unrepresented (minority class). Practical solutions using classical ML models
address the problem of large data sets using parallel/distributed
implementations of training algorithms, approximate model-based solutions, or
applying instance selection (IS) algorithms to eliminate redundant information.
However, the combined problem of big and high imbalanced datasets has been less
addressed. This work proposes three new methods for IS to be able to deal with
large and imbalanced data sets. The proposed methods use Locality Sensitive
Hashing (LSH) as a base clustering technique, and then three different sampling
methods are applied on top of the clusters (or buckets) generated by LSH. The
algorithms were developed in the Apache Spark framework, guaranteeing their
scalability. The experiments carried out in three different datasets suggest
that the proposed IS methods can improve the performance of a base ML model
between 5% and 19% in terms of the geometric mean. | {
"abstract": "Training of Machine Learning (ML) models in real contexts often deals with\nbig data sets and high-class imbalance samples where the class of interest is\nunrepresented (minority class). Practical solutions using classical ML models\naddress the problem of large data sets using parallel/distributed\nimplementations of training algorithms, approximate model-based solutions, or\napplying instance selection (IS) algorithms to eliminate redundant information.\nHowever, the combined problem of big and high imbalanced datasets has been less\naddressed. This work proposes three new methods for IS to be able to deal with\nlarge and imbalanced data sets. The proposed methods use Locality Sensitive\nHashing (LSH) as a base clustering technique, and then three different sampling\nmethods are applied on top of the clusters (or buckets) generated by LSH. The\nalgorithms were developed in the Apache Spark framework, guaranteeing their\nscalability. The experiments carried out in three different datasets suggest\nthat the proposed IS methods can improve the performance of a base ML model\nbetween 5% and 19% in terms of the geometric mean.",
"title": "An Instance Selection Algorithm for Big Data in High imbalanced datasets based on LSH",
"url": "http://arxiv.org/abs/2210.04310v1"
} | null | null | no_new_dataset | admin | null | false | null | 2587e06f-716b-4f13-9596-bfcabfb1e988 | null | Validated | {
"text_length": 1247
} | 1no_new_dataset
|
TITLE: Classification of datasets with imputed missing values: does imputation quality matter?
ABSTRACT: Classifying samples in incomplete datasets is a common aim for machine
learning practitioners, but is non-trivial. Missing data is found in most
real-world datasets and these missing values are typically imputed using
established methods, followed by classification of the now complete, imputed,
samples. The focus of the machine learning researcher is then to optimise the
downstream classification performance. In this study, we highlight that it is
imperative to consider the quality of the imputation. We demonstrate how the
commonly used measures for assessing quality are flawed and propose a new class
of discrepancy scores which focus on how well the method recreates the overall
distribution of the data. To conclude, we highlight the compromised
interpretability of classifier models trained using poorly imputed data. | {
"abstract": "Classifying samples in incomplete datasets is a common aim for machine\nlearning practitioners, but is non-trivial. Missing data is found in most\nreal-world datasets and these missing values are typically imputed using\nestablished methods, followed by classification of the now complete, imputed,\nsamples. The focus of the machine learning researcher is then to optimise the\ndownstream classification performance. In this study, we highlight that it is\nimperative to consider the quality of the imputation. We demonstrate how the\ncommonly used measures for assessing quality are flawed and propose a new class\nof discrepancy scores which focus on how well the method recreates the overall\ndistribution of the data. To conclude, we highlight the compromised\ninterpretability of classifier models trained using poorly imputed data.",
"title": "Classification of datasets with imputed missing values: does imputation quality matter?",
"url": "http://arxiv.org/abs/2206.08478v1"
} | null | null | no_new_dataset | admin | null | false | null | 2647bd44-498b-42f9-9ef8-6cb20b36eed9 | null | Validated | {
"text_length": 950
} | 1no_new_dataset
|
TITLE: Scalable mRMR feature selection to handle high dimensional datasets: Vertical partitioning based Iterative MapReduce framework
ABSTRACT: While building machine learning models, Feature selection (FS) stands out as
an essential preprocessing step used to handle the uncertainty and vagueness in
the data. Recently, the minimum Redundancy and Maximum Relevance (mRMR)
approach has proven to be effective in obtaining the irredundant feature
subset. Owing to the generation of voluminous datasets, it is essential to
design scalable solutions using distributed/parallel paradigms. MapReduce
solutions are proven to be one of the best approaches to designing
fault-tolerant and scalable solutions. This work analyses the existing
MapReduce approaches for mRMR feature selection and identifies the limitations
thereof. In the current study, we proposed VMR_mRMR, an efficient vertical
partitioning-based approach using a memorization approach, thereby overcoming
the extant approaches limitations. The experiment analysis says that VMR_mRMR
significantly outperformed extant approaches and achieved a better
computational gain (C.G). In addition, we also conducted a comparative analysis
with the horizontal partitioning approach HMR_mRMR [1] to assess the strengths
and limitations of the proposed approach. | {
"abstract": "While building machine learning models, Feature selection (FS) stands out as\nan essential preprocessing step used to handle the uncertainty and vagueness in\nthe data. Recently, the minimum Redundancy and Maximum Relevance (mRMR)\napproach has proven to be effective in obtaining the irredundant feature\nsubset. Owing to the generation of voluminous datasets, it is essential to\ndesign scalable solutions using distributed/parallel paradigms. MapReduce\nsolutions are proven to be one of the best approaches to designing\nfault-tolerant and scalable solutions. This work analyses the existing\nMapReduce approaches for mRMR feature selection and identifies the limitations\nthereof. In the current study, we proposed VMR_mRMR, an efficient vertical\npartitioning-based approach using a memorization approach, thereby overcoming\nthe extant approaches limitations. The experiment analysis says that VMR_mRMR\nsignificantly outperformed extant approaches and achieved a better\ncomputational gain (C.G). In addition, we also conducted a comparative analysis\nwith the horizontal partitioning approach HMR_mRMR [1] to assess the strengths\nand limitations of the proposed approach.",
"title": "Scalable mRMR feature selection to handle high dimensional datasets: Vertical partitioning based Iterative MapReduce framework",
"url": "http://arxiv.org/abs/2208.09901v1"
} | null | null | no_new_dataset | admin | null | false | null | 7ce72dae-c96a-4d20-b0ed-9d4e1476739b | null | Validated | {
"text_length": 1327
} | 1no_new_dataset
|
TITLE: Deep Learning based Monocular Depth Prediction: Datasets, Methods and Applications
ABSTRACT: Estimating depth from RGB images can facilitate many computer vision tasks,
such as indoor localization, height estimation, and simultaneous localization
and mapping (SLAM). Recently, monocular depth estimation has obtained great
progress owing to the rapid development of deep learning techniques. They
surpass traditional machine learning-based methods by a large margin in terms
of accuracy and speed. Despite the rapid progress in this topic, there are
lacking of a comprehensive review, which is needed to summarize the current
progress and provide the future directions. In this survey, we first introduce
the datasets for depth estimation, and then give a comprehensive introduction
of the methods from three perspectives: supervised learning-based methods,
unsupervised learning-based methods, and sparse samples guidance-based methods.
In addition, downstream applications that benefit from the progress have also
been illustrated. Finally, we point out the future directions and conclude the
paper. | {
"abstract": "Estimating depth from RGB images can facilitate many computer vision tasks,\nsuch as indoor localization, height estimation, and simultaneous localization\nand mapping (SLAM). Recently, monocular depth estimation has obtained great\nprogress owing to the rapid development of deep learning techniques. They\nsurpass traditional machine learning-based methods by a large margin in terms\nof accuracy and speed. Despite the rapid progress in this topic, there are\nlacking of a comprehensive review, which is needed to summarize the current\nprogress and provide the future directions. In this survey, we first introduce\nthe datasets for depth estimation, and then give a comprehensive introduction\nof the methods from three perspectives: supervised learning-based methods,\nunsupervised learning-based methods, and sparse samples guidance-based methods.\nIn addition, downstream applications that benefit from the progress have also\nbeen illustrated. Finally, we point out the future directions and conclude the\npaper.",
"title": "Deep Learning based Monocular Depth Prediction: Datasets, Methods and Applications",
"url": "http://arxiv.org/abs/2011.04123v1"
} | null | null | no_new_dataset | admin | null | false | null | 27cc65f3-37c2-4ed9-aa19-4a91779f034f | null | Validated | {
"text_length": 1125
} | 1no_new_dataset
|
TITLE: MRCLens: an MRC Dataset Bias Detection Toolkit
ABSTRACT: Many recent neural models have shown remarkable empirical results in Machine
Reading Comprehension, but evidence suggests sometimes the models take
advantage of dataset biases to predict and fail to generalize on out-of-sample
data. While many other approaches have been proposed to address this issue from
the computation perspective such as new architectures or training procedures,
we believe a method that allows researchers to discover biases, and adjust the
data or the models in an earlier stage will be beneficial. Thus, we introduce
MRCLens, a toolkit that detects whether biases exist before users train the
full model. For the convenience of introducing the toolkit, we also provide a
categorization of common biases in MRC. | {
"abstract": "Many recent neural models have shown remarkable empirical results in Machine\nReading Comprehension, but evidence suggests sometimes the models take\nadvantage of dataset biases to predict and fail to generalize on out-of-sample\ndata. While many other approaches have been proposed to address this issue from\nthe computation perspective such as new architectures or training procedures,\nwe believe a method that allows researchers to discover biases, and adjust the\ndata or the models in an earlier stage will be beneficial. Thus, we introduce\nMRCLens, a toolkit that detects whether biases exist before users train the\nfull model. For the convenience of introducing the toolkit, we also provide a\ncategorization of common biases in MRC.",
"title": "MRCLens: an MRC Dataset Bias Detection Toolkit",
"url": "http://arxiv.org/abs/2207.08943v1"
} | null | null | no_new_dataset | admin | null | false | null | 298e2a99-0e5b-49a9-935b-ddb37e83be36 | null | Validated | {
"text_length": 816
} | 1no_new_dataset
|
TITLE: Ex-Ante Assessment of Discrimination in Dataset
ABSTRACT: Data owners face increasing liability for how the use of their data could
harm under-priviliged communities. Stakeholders would like to identify the
characteristics of data that lead to algorithms being biased against any
particular demographic groups, for example, defined by their race, gender, age,
and/or religion. Specifically, we are interested in identifying subsets of the
feature space where the ground truth response function from features to
observed outcomes differs across demographic groups. To this end, we propose
FORESEE, a FORESt of decision trEEs algorithm, which generates a score that
captures how likely an individual's response varies with sensitive attributes.
Empirically, we find that our approach allows us to identify the individuals
who are most likely to be misclassified by several classifiers, including
Random Forest, Logistic Regression, Support Vector Machine, and k-Nearest
Neighbors. The advantage of our approach is that it allows stakeholders to
characterize risky samples that may contribute to discrimination, as well as,
use the FORESEE to estimate the risk of upcoming samples. | {
"abstract": "Data owners face increasing liability for how the use of their data could\nharm under-priviliged communities. Stakeholders would like to identify the\ncharacteristics of data that lead to algorithms being biased against any\nparticular demographic groups, for example, defined by their race, gender, age,\nand/or religion. Specifically, we are interested in identifying subsets of the\nfeature space where the ground truth response function from features to\nobserved outcomes differs across demographic groups. To this end, we propose\nFORESEE, a FORESt of decision trEEs algorithm, which generates a score that\ncaptures how likely an individual's response varies with sensitive attributes.\nEmpirically, we find that our approach allows us to identify the individuals\nwho are most likely to be misclassified by several classifiers, including\nRandom Forest, Logistic Regression, Support Vector Machine, and k-Nearest\nNeighbors. The advantage of our approach is that it allows stakeholders to\ncharacterize risky samples that may contribute to discrimination, as well as,\nuse the FORESEE to estimate the risk of upcoming samples.",
"title": "Ex-Ante Assessment of Discrimination in Dataset",
"url": "http://arxiv.org/abs/2208.07918v2"
} | null | null | no_new_dataset | admin | null | false | null | 371118da-b9ee-480b-9a75-c06b8cb1717e | null | Validated | {
"text_length": 1202
} | 1no_new_dataset
|
TITLE: Gender and Racial Bias in Visual Question Answering Datasets
ABSTRACT: Vision-and-language tasks have increasingly drawn more attention as a means
to evaluate human-like reasoning in machine learning models. A popular task in
the field is visual question answering (VQA), which aims to answer questions
about images. However, VQA models have been shown to exploit language bias by
learning the statistical correlations between questions and answers without
looking into the image content: e.g., questions about the color of a banana are
answered with yellow, even if the banana in the image is green. If societal
bias (e.g., sexism, racism, ableism, etc.) is present in the training data,
this problem may be causing VQA models to learn harmful stereotypes. For this
reason, we investigate gender and racial bias in five VQA datasets. In our
analysis, we find that the distribution of answers is highly different between
questions about women and men, as well as the existence of detrimental
gender-stereotypical samples. Likewise, we identify that specific race-related
attributes are underrepresented, whereas potentially discriminatory samples
appear in the analyzed datasets. Our findings suggest that there are dangers
associated to using VQA datasets without considering and dealing with the
potentially harmful stereotypes. We conclude the paper by proposing solutions
to alleviate the problem before, during, and after the dataset collection
process. | {
"abstract": "Vision-and-language tasks have increasingly drawn more attention as a means\nto evaluate human-like reasoning in machine learning models. A popular task in\nthe field is visual question answering (VQA), which aims to answer questions\nabout images. However, VQA models have been shown to exploit language bias by\nlearning the statistical correlations between questions and answers without\nlooking into the image content: e.g., questions about the color of a banana are\nanswered with yellow, even if the banana in the image is green. If societal\nbias (e.g., sexism, racism, ableism, etc.) is present in the training data,\nthis problem may be causing VQA models to learn harmful stereotypes. For this\nreason, we investigate gender and racial bias in five VQA datasets. In our\nanalysis, we find that the distribution of answers is highly different between\nquestions about women and men, as well as the existence of detrimental\ngender-stereotypical samples. Likewise, we identify that specific race-related\nattributes are underrepresented, whereas potentially discriminatory samples\nappear in the analyzed datasets. Our findings suggest that there are dangers\nassociated to using VQA datasets without considering and dealing with the\npotentially harmful stereotypes. We conclude the paper by proposing solutions\nto alleviate the problem before, during, and after the dataset collection\nprocess.",
"title": "Gender and Racial Bias in Visual Question Answering Datasets",
"url": "http://arxiv.org/abs/2205.08148v3"
} | null | null | no_new_dataset | admin | null | false | null | a2dc7836-2bbd-4b76-b785-1e06276846c0 | null | Validated | {
"text_length": 1482
} | 1no_new_dataset
|
TITLE: Computer Vision based inspection on post-earthquake with UAV synthetic dataset
ABSTRACT: The area affected by the earthquake is vast and often difficult to entirely
cover, and the earthquake itself is a sudden event that causes multiple defects
simultaneously, that cannot be effectively traced using traditional, manual
methods. This article presents an innovative approach to the problem of
detecting damage after sudden events by using an interconnected set of deep
machine learning models organized in a single pipeline and allowing for easy
modification and swapping models seamlessly. Models in the pipeline were
trained with a synthetic dataset and were adapted to be further evaluated and
used with unmanned aerial vehicles (UAVs) in real-world conditions. Thanks to
the methods presented in the article, it is possible to obtain high accuracy in
detecting buildings defects, segmenting constructions into their components and
estimating their technical condition based on a single drone flight. | {
"abstract": "The area affected by the earthquake is vast and often difficult to entirely\ncover, and the earthquake itself is a sudden event that causes multiple defects\nsimultaneously, that cannot be effectively traced using traditional, manual\nmethods. This article presents an innovative approach to the problem of\ndetecting damage after sudden events by using an interconnected set of deep\nmachine learning models organized in a single pipeline and allowing for easy\nmodification and swapping models seamlessly. Models in the pipeline were\ntrained with a synthetic dataset and were adapted to be further evaluated and\nused with unmanned aerial vehicles (UAVs) in real-world conditions. Thanks to\nthe methods presented in the article, it is possible to obtain high accuracy in\ndetecting buildings defects, segmenting constructions into their components and\nestimating their technical condition based on a single drone flight.",
"title": "Computer Vision based inspection on post-earthquake with UAV synthetic dataset",
"url": "http://arxiv.org/abs/2210.05282v1"
} | null | null | no_new_dataset | admin | null | false | null | 436923c6-aa42-45c0-a8a4-7e6e9dbb60e4 | null | Validated | {
"text_length": 1027
} | 1no_new_dataset
|
TITLE: HAPI: A Large-scale Longitudinal Dataset of Commercial ML API Predictions
ABSTRACT: Commercial ML APIs offered by providers such as Google, Amazon and Microsoft
have dramatically simplified ML adoption in many applications. Numerous
companies and academics pay to use ML APIs for tasks such as object detection,
OCR and sentiment analysis. Different ML APIs tackling the same task can have
very heterogeneous performance. Moreover, the ML models underlying the APIs
also evolve over time. As ML APIs rapidly become a valuable marketplace and a
widespread way to consume machine learning, it is critical to systematically
study and compare different APIs with each other and to characterize how APIs
change over time. However, this topic is currently underexplored due to the
lack of data. In this paper, we present HAPI (History of APIs), a longitudinal
dataset of 1,761,417 instances of commercial ML API applications (involving
APIs from Amazon, Google, IBM, Microsoft and other providers) across diverse
tasks including image tagging, speech recognition and text mining from 2020 to
2022. Each instance consists of a query input for an API (e.g., an image or
text) along with the API's output prediction/annotation and confidence scores.
HAPI is the first large-scale dataset of ML API usages and is a unique resource
for studying ML-as-a-service (MLaaS). As examples of the types of analyses that
HAPI enables, we show that ML APIs' performance change substantially over
time--several APIs' accuracies dropped on specific benchmark datasets. Even
when the API's aggregate performance stays steady, its error modes can shift
across different subtypes of data between 2020 and 2022. Such changes can
substantially impact the entire analytics pipelines that use some ML API as a
component. We further use HAPI to study commercial APIs' performance
disparities across demographic subgroups over time. HAPI can stimulate more
research in the growing field of MLaaS. | {
"abstract": "Commercial ML APIs offered by providers such as Google, Amazon and Microsoft\nhave dramatically simplified ML adoption in many applications. Numerous\ncompanies and academics pay to use ML APIs for tasks such as object detection,\nOCR and sentiment analysis. Different ML APIs tackling the same task can have\nvery heterogeneous performance. Moreover, the ML models underlying the APIs\nalso evolve over time. As ML APIs rapidly become a valuable marketplace and a\nwidespread way to consume machine learning, it is critical to systematically\nstudy and compare different APIs with each other and to characterize how APIs\nchange over time. However, this topic is currently underexplored due to the\nlack of data. In this paper, we present HAPI (History of APIs), a longitudinal\ndataset of 1,761,417 instances of commercial ML API applications (involving\nAPIs from Amazon, Google, IBM, Microsoft and other providers) across diverse\ntasks including image tagging, speech recognition and text mining from 2020 to\n2022. Each instance consists of a query input for an API (e.g., an image or\ntext) along with the API's output prediction/annotation and confidence scores.\nHAPI is the first large-scale dataset of ML API usages and is a unique resource\nfor studying ML-as-a-service (MLaaS). As examples of the types of analyses that\nHAPI enables, we show that ML APIs' performance change substantially over\ntime--several APIs' accuracies dropped on specific benchmark datasets. Even\nwhen the API's aggregate performance stays steady, its error modes can shift\nacross different subtypes of data between 2020 and 2022. Such changes can\nsubstantially impact the entire analytics pipelines that use some ML API as a\ncomponent. We further use HAPI to study commercial APIs' performance\ndisparities across demographic subgroups over time. HAPI can stimulate more\nresearch in the growing field of MLaaS.",
"title": "HAPI: A Large-scale Longitudinal Dataset of Commercial ML API Predictions",
"url": "http://arxiv.org/abs/2209.08443v1"
} | null | null | new_dataset | admin | null | false | null | dbb8c2fe-bb43-4e45-9fd7-15a622ece478 | null | Validated | {
"text_length": 1988
} | 0new_dataset
|
TITLE: GLARE: A Dataset for Traffic Sign Detection in Sun Glare
ABSTRACT: Real-time machine learning detection algorithms are often found within
autonomous vehicle technology and depend on quality datasets. It is essential
that these algorithms work correctly in everyday conditions as well as under
strong sun glare. Reports indicate glare is one of the two most prominent
environment-related reasons for crashes. However, existing datasets, such as
LISA and the German Traffic Sign Recognition Benchmark, do not reflect the
existence of sun glare at all. This paper presents the GLARE traffic sign
dataset: a collection of images with U.S based traffic signs under heavy visual
interference by sunlight. GLARE contains 2,157 images of traffic signs with sun
glare, pulled from 33 videos of dashcam footage of roads in the United States.
It provides an essential enrichment to the widely used LISA Traffic Sign
dataset. Our experimental study shows that although several state-of-the-art
baseline methods demonstrate superior performance when trained and tested
against traffic sign datasets without sun glare, they greatly suffer when
tested against GLARE (e.g., ranging from 9% to 21% mean mAP, which is
significantly lower than the performances on LISA dataset). We also notice that
current architectures have better detection accuracy (e.g., on average 42% mean
mAP gain for mainstream algorithms) when trained on images of traffic signs in
sun glare. | {
"abstract": "Real-time machine learning detection algorithms are often found within\nautonomous vehicle technology and depend on quality datasets. It is essential\nthat these algorithms work correctly in everyday conditions as well as under\nstrong sun glare. Reports indicate glare is one of the two most prominent\nenvironment-related reasons for crashes. However, existing datasets, such as\nLISA and the German Traffic Sign Recognition Benchmark, do not reflect the\nexistence of sun glare at all. This paper presents the GLARE traffic sign\ndataset: a collection of images with U.S based traffic signs under heavy visual\ninterference by sunlight. GLARE contains 2,157 images of traffic signs with sun\nglare, pulled from 33 videos of dashcam footage of roads in the United States.\nIt provides an essential enrichment to the widely used LISA Traffic Sign\ndataset. Our experimental study shows that although several state-of-the-art\nbaseline methods demonstrate superior performance when trained and tested\nagainst traffic sign datasets without sun glare, they greatly suffer when\ntested against GLARE (e.g., ranging from 9% to 21% mean mAP, which is\nsignificantly lower than the performances on LISA dataset). We also notice that\ncurrent architectures have better detection accuracy (e.g., on average 42% mean\nmAP gain for mainstream algorithms) when trained on images of traffic signs in\nsun glare.",
"title": "GLARE: A Dataset for Traffic Sign Detection in Sun Glare",
"url": "http://arxiv.org/abs/2209.08716v1"
} | null | null | new_dataset | admin | null | false | null | 8a57c963-93a0-4088-8f7b-93262692f72a | null | Validated | {
"text_length": 1473
} | 0new_dataset
|
TITLE: COVIDx CT-3: A Large-scale, Multinational, Open-Source Benchmark Dataset for Computer-aided COVID-19 Screening from Chest CT Images
ABSTRACT: Computed tomography (CT) has been widely explored as a COVID-19 screening and
assessment tool to complement RT-PCR testing. To assist radiologists with
CT-based COVID-19 screening, a number of computer-aided systems have been
proposed. However, many proposed systems are built using CT data which is
limited in both quantity and diversity. Motivated to support efforts in the
development of machine learning-driven screening systems, we introduce COVIDx
CT-3, a large-scale multinational benchmark dataset for detection of COVID-19
cases from chest CT images. COVIDx CT-3 includes 431,205 CT slices from 6,068
patients across at least 17 countries, which to the best of our knowledge
represents the largest, most diverse dataset of COVID-19 CT images in
open-access form. Additionally, we examine the data diversity and potential
biases of the COVIDx CT-3 dataset, finding that significant geographic and
class imbalances remain despite efforts to curate data from a wide variety of
sources. | {
"abstract": "Computed tomography (CT) has been widely explored as a COVID-19 screening and\nassessment tool to complement RT-PCR testing. To assist radiologists with\nCT-based COVID-19 screening, a number of computer-aided systems have been\nproposed. However, many proposed systems are built using CT data which is\nlimited in both quantity and diversity. Motivated to support efforts in the\ndevelopment of machine learning-driven screening systems, we introduce COVIDx\nCT-3, a large-scale multinational benchmark dataset for detection of COVID-19\ncases from chest CT images. COVIDx CT-3 includes 431,205 CT slices from 6,068\npatients across at least 17 countries, which to the best of our knowledge\nrepresents the largest, most diverse dataset of COVID-19 CT images in\nopen-access form. Additionally, we examine the data diversity and potential\nbiases of the COVIDx CT-3 dataset, finding that significant geographic and\nclass imbalances remain despite efforts to curate data from a wide variety of\nsources.",
"title": "COVIDx CT-3: A Large-scale, Multinational, Open-Source Benchmark Dataset for Computer-aided COVID-19 Screening from Chest CT Images",
"url": "http://arxiv.org/abs/2206.03043v3"
} | null | null | new_dataset | admin | null | false | null | 14f90c95-ff42-402d-8bb3-ccd3ece8cc0d | null | Validated | {
"text_length": 1157
} | 0new_dataset
|
TITLE: Shifts 2.0: Extending The Dataset of Real Distributional Shifts
ABSTRACT: Distributional shift, or the mismatch between training and deployment data,
is a significant obstacle to the usage of machine learning in high-stakes
industrial applications, such as autonomous driving and medicine. This creates
a need to be able to assess how robustly ML models generalize as well as the
quality of their uncertainty estimates. Standard ML baseline datasets do not
allow these properties to be assessed, as the training, validation and test
data are often identically distributed. Recently, a range of dedicated
benchmarks have appeared, featuring both distributionally matched and shifted
data. Among these benchmarks, the Shifts dataset stands out in terms of the
diversity of tasks as well as the data modalities it features. While most of
the benchmarks are heavily dominated by 2D image classification tasks, Shifts
contains tabular weather forecasting, machine translation, and vehicle motion
prediction tasks. This enables the robustness properties of models to be
assessed on a diverse set of industrial-scale tasks and either universal or
directly applicable task-specific conclusions to be reached. In this paper, we
extend the Shifts Dataset with two datasets sourced from industrial, high-risk
applications of high societal importance. Specifically, we consider the tasks
of segmentation of white matter Multiple Sclerosis lesions in 3D magnetic
resonance brain images and the estimation of power consumption in marine cargo
vessels. Both tasks feature ubiquitous distributional shifts and a strict
safety requirement due to the high cost of errors. These new datasets will
allow researchers to further explore robust generalization and uncertainty
estimation in new situations. In this work, we provide a description of the
dataset and baseline results for both tasks. | {
"abstract": "Distributional shift, or the mismatch between training and deployment data,\nis a significant obstacle to the usage of machine learning in high-stakes\nindustrial applications, such as autonomous driving and medicine. This creates\na need to be able to assess how robustly ML models generalize as well as the\nquality of their uncertainty estimates. Standard ML baseline datasets do not\nallow these properties to be assessed, as the training, validation and test\ndata are often identically distributed. Recently, a range of dedicated\nbenchmarks have appeared, featuring both distributionally matched and shifted\ndata. Among these benchmarks, the Shifts dataset stands out in terms of the\ndiversity of tasks as well as the data modalities it features. While most of\nthe benchmarks are heavily dominated by 2D image classification tasks, Shifts\ncontains tabular weather forecasting, machine translation, and vehicle motion\nprediction tasks. This enables the robustness properties of models to be\nassessed on a diverse set of industrial-scale tasks and either universal or\ndirectly applicable task-specific conclusions to be reached. In this paper, we\nextend the Shifts Dataset with two datasets sourced from industrial, high-risk\napplications of high societal importance. Specifically, we consider the tasks\nof segmentation of white matter Multiple Sclerosis lesions in 3D magnetic\nresonance brain images and the estimation of power consumption in marine cargo\nvessels. Both tasks feature ubiquitous distributional shifts and a strict\nsafety requirement due to the high cost of errors. These new datasets will\nallow researchers to further explore robust generalization and uncertainty\nestimation in new situations. In this work, we provide a description of the\ndataset and baseline results for both tasks.",
"title": "Shifts 2.0: Extending The Dataset of Real Distributional Shifts",
"url": "http://arxiv.org/abs/2206.15407v2"
} | null | null | new_dataset | admin | null | false | null | bb2169db-72b3-4ce2-85ae-d66da1e2fd03 | null | Validated | {
"text_length": 1897
} | 0new_dataset
|
TITLE: Machine Learning Performance Analysis to Predict Stroke Based on Imbalanced Medical Dataset
ABSTRACT: Cerebral stroke, the second most substantial cause of death universally, has
been a primary public health concern over the last few years. With the help of
machine learning techniques, early detection of various stroke alerts is
accessible, which can efficiently prevent or diminish the stroke. Medical
dataset, however, are frequently unbalanced in their class label, with a
tendency to poorly predict minority classes. In this paper, the potential risk
factors for stroke are investigated. Moreover, four distinctive approaches are
applied to improve the classification of the minority class in the imbalanced
stroke dataset, which are the ensemble weight voting classifier, the Synthetic
Minority Over-sampling Technique (SMOTE), Principal Component Analysis with
K-Means Clustering (PCA-Kmeans), Focal Loss with the Deep Neural Network (DNN)
and compare their performance. Through the analysis results, SMOTE and
PCA-Kmeans with DNN-Focal Loss work best for the limited size of a large severe
imbalanced dataset,which is 2-4 times outperform Kaggle work. | {
"abstract": "Cerebral stroke, the second most substantial cause of death universally, has\nbeen a primary public health concern over the last few years. With the help of\nmachine learning techniques, early detection of various stroke alerts is\naccessible, which can efficiently prevent or diminish the stroke. Medical\ndataset, however, are frequently unbalanced in their class label, with a\ntendency to poorly predict minority classes. In this paper, the potential risk\nfactors for stroke are investigated. Moreover, four distinctive approaches are\napplied to improve the classification of the minority class in the imbalanced\nstroke dataset, which are the ensemble weight voting classifier, the Synthetic\nMinority Over-sampling Technique (SMOTE), Principal Component Analysis with\nK-Means Clustering (PCA-Kmeans), Focal Loss with the Deep Neural Network (DNN)\nand compare their performance. Through the analysis results, SMOTE and\nPCA-Kmeans with DNN-Focal Loss work best for the limited size of a large severe\nimbalanced dataset,which is 2-4 times outperform Kaggle work.",
"title": "Machine Learning Performance Analysis to Predict Stroke Based on Imbalanced Medical Dataset",
"url": "http://arxiv.org/abs/2211.07652v1"
} | null | null | no_new_dataset | admin | null | false | null | 935e37df-adc5-49be-835a-24ac558f5c98 | null | Validated | {
"text_length": 1184
} | 1no_new_dataset
|
TITLE: Evaluating resampling methods on a real-life highly imbalanced online credit card payments dataset
ABSTRACT: Various problems of any credit card fraud detection based on machine learning
come from the imbalanced aspect of transaction datasets. Indeed, the number of
frauds compared to the number of regular transactions is tiny and has been
shown to damage learning performances, e.g., at worst, the algorithm can learn
to classify all the transactions as regular. Resampling methods and
cost-sensitive approaches are known to be good candidates to leverage this
issue of imbalanced datasets. This paper evaluates numerous state-of-the-art
resampling methods on a large real-life online credit card payments dataset. We
show they are inefficient because methods are intractable or because metrics do
not exhibit substantial improvements. Our work contributes to this domain in
(1) that we compare many state-of-the-art resampling methods on a large-scale
dataset and in (2) that we use a real-life online credit card payments dataset. | {
"abstract": "Various problems of any credit card fraud detection based on machine learning\ncome from the imbalanced aspect of transaction datasets. Indeed, the number of\nfrauds compared to the number of regular transactions is tiny and has been\nshown to damage learning performances, e.g., at worst, the algorithm can learn\nto classify all the transactions as regular. Resampling methods and\ncost-sensitive approaches are known to be good candidates to leverage this\nissue of imbalanced datasets. This paper evaluates numerous state-of-the-art\nresampling methods on a large real-life online credit card payments dataset. We\nshow they are inefficient because methods are intractable or because metrics do\nnot exhibit substantial improvements. Our work contributes to this domain in\n(1) that we compare many state-of-the-art resampling methods on a large-scale\ndataset and in (2) that we use a real-life online credit card payments dataset.",
"title": "Evaluating resampling methods on a real-life highly imbalanced online credit card payments dataset",
"url": "http://arxiv.org/abs/2206.13152v1"
} | null | null | no_new_dataset | admin | null | false | null | e32c31dc-e430-4cce-bda1-78b8ad39277f | null | Default | {
"text_length": 1058
} | 1no_new_dataset
|
TITLE: The Conversational Short-phrase Speaker Diarization (CSSD) Task: Dataset, Evaluation Metric and Baselines
ABSTRACT: The conversation scenario is one of the most important and most challenging
scenarios for speech processing technologies because people in conversation
respond to each other in a casual style. Detecting the speech activities of
each person in a conversation is vital to downstream tasks, like natural
language processing, machine translation, etc. People refer to the detection
technology of "who speak when" as speaker diarization (SD). Traditionally,
diarization error rate (DER) has been used as the standard evaluation metric of
SD systems for a long time. However, DER fails to give enough importance to
short conversational phrases, which are short but important on the semantic
level. Also, a carefully and accurately manually-annotated testing dataset
suitable for evaluating the conversational SD technologies is still unavailable
in the speech community. In this paper, we design and describe the
Conversational Short-phrases Speaker Diarization (CSSD) task, which consists of
training and testing datasets, evaluation metric and baselines. In the dataset
aspect, despite the previously open-sourced 180-hour conversational
MagicData-RAMC dataset, we prepare an individual 20-hour conversational speech
test dataset with carefully and artificially verified speakers timestamps
annotations for the CSSD task. In the metric aspect, we design the new
conversational DER (CDER) evaluation metric, which calculates the SD accuracy
at the utterance level. In the baseline aspect, we adopt a commonly used
method: Variational Bayes HMM x-vector system, as the baseline of the CSSD
task. Our evaluation metric is publicly available at
https://github.com/SpeechClub/CDER_Metric. | {
"abstract": "The conversation scenario is one of the most important and most challenging\nscenarios for speech processing technologies because people in conversation\nrespond to each other in a casual style. Detecting the speech activities of\neach person in a conversation is vital to downstream tasks, like natural\nlanguage processing, machine translation, etc. People refer to the detection\ntechnology of \"who speak when\" as speaker diarization (SD). Traditionally,\ndiarization error rate (DER) has been used as the standard evaluation metric of\nSD systems for a long time. However, DER fails to give enough importance to\nshort conversational phrases, which are short but important on the semantic\nlevel. Also, a carefully and accurately manually-annotated testing dataset\nsuitable for evaluating the conversational SD technologies is still unavailable\nin the speech community. In this paper, we design and describe the\nConversational Short-phrases Speaker Diarization (CSSD) task, which consists of\ntraining and testing datasets, evaluation metric and baselines. In the dataset\naspect, despite the previously open-sourced 180-hour conversational\nMagicData-RAMC dataset, we prepare an individual 20-hour conversational speech\ntest dataset with carefully and artificially verified speakers timestamps\nannotations for the CSSD task. In the metric aspect, we design the new\nconversational DER (CDER) evaluation metric, which calculates the SD accuracy\nat the utterance level. In the baseline aspect, we adopt a commonly used\nmethod: Variational Bayes HMM x-vector system, as the baseline of the CSSD\ntask. Our evaluation metric is publicly available at\nhttps://github.com/SpeechClub/CDER_Metric.",
"title": "The Conversational Short-phrase Speaker Diarization (CSSD) Task: Dataset, Evaluation Metric and Baselines",
"url": "http://arxiv.org/abs/2208.08042v1"
} | null | null | new_dataset | admin | null | false | null | fd49659a-52a0-4e86-a6b0-5c2555520522 | null | Validated | {
"text_length": 1819
} | 0new_dataset
|
TITLE: Solar Active Region Magnetogram Image Dataset for Studies of Space Weather
ABSTRACT: In this dataset we provide a comprehensive collection of magnetograms (images
quantifying the strength of the magnetic field) from the National Aeronautics
and Space Administration's (NASA's) Solar Dynamics Observatory (SDO). The
dataset incorporates data from three sources and provides SDO Helioseismic and
Magnetic Imager (HMI) magnetograms of solar active regions (regions of large
magnetic flux, generally the source of eruptive events) as well as labels of
corresponding flaring activity. This dataset will be useful for image analysis
or solar physics research related to magnetic structure, its evolution over
time, and its relation to solar flares. The dataset will be of interest to
those researchers investigating automated solar flare prediction methods,
including supervised and unsupervised machine learning (classical and deep),
binary and multi-class classification, and regression. This dataset is a
minimally processed, user configurable dataset of consistently sized images of
solar active regions that can serve as a benchmark dataset for solar flare
prediction research. | {
"abstract": "In this dataset we provide a comprehensive collection of magnetograms (images\nquantifying the strength of the magnetic field) from the National Aeronautics\nand Space Administration's (NASA's) Solar Dynamics Observatory (SDO). The\ndataset incorporates data from three sources and provides SDO Helioseismic and\nMagnetic Imager (HMI) magnetograms of solar active regions (regions of large\nmagnetic flux, generally the source of eruptive events) as well as labels of\ncorresponding flaring activity. This dataset will be useful for image analysis\nor solar physics research related to magnetic structure, its evolution over\ntime, and its relation to solar flares. The dataset will be of interest to\nthose researchers investigating automated solar flare prediction methods,\nincluding supervised and unsupervised machine learning (classical and deep),\nbinary and multi-class classification, and regression. This dataset is a\nminimally processed, user configurable dataset of consistently sized images of\nsolar active regions that can serve as a benchmark dataset for solar flare\nprediction research.",
"title": "Solar Active Region Magnetogram Image Dataset for Studies of Space Weather",
"url": "http://arxiv.org/abs/2305.09492v2"
} | null | null | new_dataset | admin | null | false | null | eb0b9191-def7-4bf5-b92f-74c249692b45 | null | Validated | {
"text_length": 1200
} | 0new_dataset
|
TITLE: Commander's Intent: A Dataset and Modeling Approach for Human-AI Task Specification in Strategic Play
ABSTRACT: Effective Human-AI teaming requires the ability to communicate the goals of
the team and constraints under which you need the agent to operate. Providing
the ability to specify the shared intent or operation criteria of the team can
enable an AI agent to perform its primary function while still being able to
cater to the specific desires of the current team. While significant work has
been conducted to instruct an agent to perform a task, via language or
demonstrations, prior work lacks a focus on building agents which can operate
within the parameters specified by a team. Worse yet, there is a dearth of
research pertaining to enabling humans to provide their specifications through
unstructured, naturalist language. In this paper, we propose the use of goals
and constraints as a scaffold to modulate and evaluate autonomous agents. We
contribute to this field by presenting a novel dataset, and an associated data
collection protocol, which maps language descriptions to goals and constraints
corresponding to specific strategies developed by human participants for the
board game Risk. Leveraging state-of-the-art language models and augmentation
procedures, we develop a machine learning framework which can be used to
identify goals and constraints from unstructured strategy descriptions. To
empirically validate our approach we conduct a human-subjects study to
establish a human-baseline for our dataset. Our results show that our machine
learning architecture is better able to interpret unstructured language
descriptions into strategy specifications than human raters tasked with
performing the same machine translation task (F(1,272.53) = 17.025, p < 0.001). | {
"abstract": "Effective Human-AI teaming requires the ability to communicate the goals of\nthe team and constraints under which you need the agent to operate. Providing\nthe ability to specify the shared intent or operation criteria of the team can\nenable an AI agent to perform its primary function while still being able to\ncater to the specific desires of the current team. While significant work has\nbeen conducted to instruct an agent to perform a task, via language or\ndemonstrations, prior work lacks a focus on building agents which can operate\nwithin the parameters specified by a team. Worse yet, there is a dearth of\nresearch pertaining to enabling humans to provide their specifications through\nunstructured, naturalist language. In this paper, we propose the use of goals\nand constraints as a scaffold to modulate and evaluate autonomous agents. We\ncontribute to this field by presenting a novel dataset, and an associated data\ncollection protocol, which maps language descriptions to goals and constraints\ncorresponding to specific strategies developed by human participants for the\nboard game Risk. Leveraging state-of-the-art language models and augmentation\nprocedures, we develop a machine learning framework which can be used to\nidentify goals and constraints from unstructured strategy descriptions. To\nempirically validate our approach we conduct a human-subjects study to\nestablish a human-baseline for our dataset. Our results show that our machine\nlearning architecture is better able to interpret unstructured language\ndescriptions into strategy specifications than human raters tasked with\nperforming the same machine translation task (F(1,272.53) = 17.025, p < 0.001).",
"title": "Commander's Intent: A Dataset and Modeling Approach for Human-AI Task Specification in Strategic Play",
"url": "http://arxiv.org/abs/2208.08374v1"
} | null | null | new_dataset | admin | null | false | null | f59375fa-83ae-406a-b541-3df9881bafce | null | Validated | {
"text_length": 1815
} | 0new_dataset
|
TITLE: TruEyes: Utilizing Microtasks in Mobile Apps for Crowdsourced Labeling of Machine Learning Datasets
ABSTRACT: The growing use of supervised machine learning in research and industry has
increased the need for labeled datasets. Crowdsourcing has emerged as a popular
method to create data labels. However, working on large batches of tasks leads
to worker fatigue, negatively impacting labeling quality. To address this, we
present TruEyes, a collaborative crowdsourcing system, enabling the
distribution of micro-tasks to mobile app users. TruEyes allows machine
learning practitioners to publish labeling tasks, mobile app developers to
integrate task ads for monetization, and users to label data instead of
watching advertisements. To evaluate the system, we conducted an experiment
with N=296 participants. Our results show that the quality of the labeled data
is comparable to traditional crowdsourcing approaches and most users prefer
task ads over traditional ads. We discuss extensions to the system and address
how mobile advertisement space can be used as a productive resource in the
future. | {
"abstract": "The growing use of supervised machine learning in research and industry has\nincreased the need for labeled datasets. Crowdsourcing has emerged as a popular\nmethod to create data labels. However, working on large batches of tasks leads\nto worker fatigue, negatively impacting labeling quality. To address this, we\npresent TruEyes, a collaborative crowdsourcing system, enabling the\ndistribution of micro-tasks to mobile app users. TruEyes allows machine\nlearning practitioners to publish labeling tasks, mobile app developers to\nintegrate task ads for monetization, and users to label data instead of\nwatching advertisements. To evaluate the system, we conducted an experiment\nwith N=296 participants. Our results show that the quality of the labeled data\nis comparable to traditional crowdsourcing approaches and most users prefer\ntask ads over traditional ads. We discuss extensions to the system and address\nhow mobile advertisement space can be used as a productive resource in the\nfuture.",
"title": "TruEyes: Utilizing Microtasks in Mobile Apps for Crowdsourced Labeling of Machine Learning Datasets",
"url": "http://arxiv.org/abs/2209.14708v1"
} | null | null | no_new_dataset | admin | null | false | null | 72f7f796-bceb-44cc-a1c2-93ddbc5e027c | null | Validated | {
"text_length": 1126
} | 1no_new_dataset
|
TITLE: Data Models for Dataset Drift Controls in Machine Learning With Optical Images
ABSTRACT: Camera images are ubiquitous in machine learning research. They also play a
central role in the delivery of important services spanning medicine and
environmental surveying. However, the application of machine learning models in
these domains has been limited because of robustness concerns. A primary
failure mode are performance drops due to differences between the training and
deployment data. While there are methods to prospectively validate the
robustness of machine learning models to such dataset drifts, existing
approaches do not account for explicit models of the primary object of
interest: the data. This limits our ability to study and understand the
relationship between data generation and downstream machine learning model
performance in a physically accurate manner. In this study, we demonstrate how
to overcome this limitation by pairing traditional machine learning with
physical optics to obtain explicit and differentiable data models. We
demonstrate how such data models can be constructed for image data and used to
control downstream machine learning model performance related to dataset drift.
The findings are distilled into three applications. First, drift synthesis
enables the controlled generation of physically faithful drift test cases to
power model selection and targeted generalization. Second, the gradient
connection between machine learning task model and data model allows advanced,
precise tolerancing of task model sensitivity to changes in the data
generation. These drift forensics can be used to precisely specify the
acceptable data environments in which a task model may be run. Third, drift
optimization opens up the possibility to create drifts that can help the task
model learn better faster, effectively optimizing the data generating process
itself. A guide to access the open code and datasets is available at
https://github.com/aiaudit-org/raw2logit. | {
"abstract": "Camera images are ubiquitous in machine learning research. They also play a\ncentral role in the delivery of important services spanning medicine and\nenvironmental surveying. However, the application of machine learning models in\nthese domains has been limited because of robustness concerns. A primary\nfailure mode are performance drops due to differences between the training and\ndeployment data. While there are methods to prospectively validate the\nrobustness of machine learning models to such dataset drifts, existing\napproaches do not account for explicit models of the primary object of\ninterest: the data. This limits our ability to study and understand the\nrelationship between data generation and downstream machine learning model\nperformance in a physically accurate manner. In this study, we demonstrate how\nto overcome this limitation by pairing traditional machine learning with\nphysical optics to obtain explicit and differentiable data models. We\ndemonstrate how such data models can be constructed for image data and used to\ncontrol downstream machine learning model performance related to dataset drift.\nThe findings are distilled into three applications. First, drift synthesis\nenables the controlled generation of physically faithful drift test cases to\npower model selection and targeted generalization. Second, the gradient\nconnection between machine learning task model and data model allows advanced,\nprecise tolerancing of task model sensitivity to changes in the data\ngeneration. These drift forensics can be used to precisely specify the\nacceptable data environments in which a task model may be run. Third, drift\noptimization opens up the possibility to create drifts that can help the task\nmodel learn better faster, effectively optimizing the data generating process\nitself. A guide to access the open code and datasets is available at\nhttps://github.com/aiaudit-org/raw2logit.",
"title": "Data Models for Dataset Drift Controls in Machine Learning With Optical Images",
"url": "http://arxiv.org/abs/2211.02578v3"
} | null | null | no_new_dataset | admin | null | false | null | 75bfa54b-db90-45f0-89a9-3102ff3b3df3 | null | Validated | {
"text_length": 2020
} | 1no_new_dataset
|
TITLE: The ITU Faroese Pairs Dataset
ABSTRACT: This article documents a dataset of sentence pairs between Faroese and
Danish, produced at ITU Copenhagen. The data covers tranlsation from both
source languages, and is intended for use as training data for machine
translation systems in this language pair. | {
"abstract": "This article documents a dataset of sentence pairs between Faroese and\nDanish, produced at ITU Copenhagen. The data covers tranlsation from both\nsource languages, and is intended for use as training data for machine\ntranslation systems in this language pair.",
"title": "The ITU Faroese Pairs Dataset",
"url": "http://arxiv.org/abs/2206.08727v1"
} | null | null | new_dataset | admin | null | false | null | 2f29679a-049c-4cee-b4d7-97905b3c9b6c | null | Validated | {
"text_length": 322
} | 0new_dataset
|
TITLE: Improving Multilayer-Perceptron(MLP)-based Network Anomaly Detection with Birch Clustering on CICIDS-2017 Dataset
ABSTRACT: Machine learning algorithms have been widely used in intrusion detection
systems, including Multi-layer Perceptron (MLP). In this study, we proposed a
two-stage model that combines the Birch clustering algorithm and MLP classifier
to improve the performance of network anomaly multi-classification. In our
proposed method, we first apply Birch or Kmeans as an unsupervised clustering
algorithm to the CICIDS-2017 dataset to pre-group the data. The generated
pseudo-label is then added as an additional feature to the training of the
MLP-based classifier. The experimental results show that using Birch and
K-Means clustering for data pre-grouping can improve intrusion detection system
performance. Our method can achieve 99.73% accuracy in multi-classification
using Birch clustering, which is better than similar researches using a
stand-alone MLP model. | {
"abstract": "Machine learning algorithms have been widely used in intrusion detection\nsystems, including Multi-layer Perceptron (MLP). In this study, we proposed a\ntwo-stage model that combines the Birch clustering algorithm and MLP classifier\nto improve the performance of network anomaly multi-classification. In our\nproposed method, we first apply Birch or Kmeans as an unsupervised clustering\nalgorithm to the CICIDS-2017 dataset to pre-group the data. The generated\npseudo-label is then added as an additional feature to the training of the\nMLP-based classifier. The experimental results show that using Birch and\nK-Means clustering for data pre-grouping can improve intrusion detection system\nperformance. Our method can achieve 99.73% accuracy in multi-classification\nusing Birch clustering, which is better than similar researches using a\nstand-alone MLP model.",
"title": "Improving Multilayer-Perceptron(MLP)-based Network Anomaly Detection with Birch Clustering on CICIDS-2017 Dataset",
"url": "http://arxiv.org/abs/2208.09711v2"
} | null | null | no_new_dataset | admin | null | false | null | 60902729-ac8e-43f0-b661-6bdcc36c3f92 | null | Validated | {
"text_length": 1004
} | 1no_new_dataset
|
TITLE: PcMSP: A Dataset for Scientific Action Graphs Extraction from Polycrystalline Materials Synthesis Procedure Text
ABSTRACT: Scientific action graphs extraction from materials synthesis procedures is
important for reproducible research, machine automation, and material
prediction. But the lack of annotated data has hindered progress in this field.
We demonstrate an effort to annotate Polycrystalline Materials Synthesis
Procedures (PcMSP) from 305 open access scientific articles for the
construction of synthesis action graphs. This is a new dataset for material
science information extraction that simultaneously contains the synthesis
sentences extracted from the experimental paragraphs, as well as the entity
mentions and intra-sentence relations. A two-step human annotation and
inter-annotator agreement study guarantee the high quality of the PcMSP corpus.
We introduce four natural language processing tasks: sentence classification,
named entity recognition, relation classification, and joint extraction of
entities and relations. Comprehensive experiments validate the effectiveness of
several state-of-the-art models for these challenges while leaving large space
for improvement. We also perform the error analysis and point out some unique
challenges that require further investigation. We will release our annotation
scheme, the corpus, and codes to the research community to alleviate the
scarcity of labeled data in this domain. | {
"abstract": "Scientific action graphs extraction from materials synthesis procedures is\nimportant for reproducible research, machine automation, and material\nprediction. But the lack of annotated data has hindered progress in this field.\nWe demonstrate an effort to annotate Polycrystalline Materials Synthesis\nProcedures (PcMSP) from 305 open access scientific articles for the\nconstruction of synthesis action graphs. This is a new dataset for material\nscience information extraction that simultaneously contains the synthesis\nsentences extracted from the experimental paragraphs, as well as the entity\nmentions and intra-sentence relations. A two-step human annotation and\ninter-annotator agreement study guarantee the high quality of the PcMSP corpus.\nWe introduce four natural language processing tasks: sentence classification,\nnamed entity recognition, relation classification, and joint extraction of\nentities and relations. Comprehensive experiments validate the effectiveness of\nseveral state-of-the-art models for these challenges while leaving large space\nfor improvement. We also perform the error analysis and point out some unique\nchallenges that require further investigation. We will release our annotation\nscheme, the corpus, and codes to the research community to alleviate the\nscarcity of labeled data in this domain.",
"title": "PcMSP: A Dataset for Scientific Action Graphs Extraction from Polycrystalline Materials Synthesis Procedure Text",
"url": "http://arxiv.org/abs/2210.12401v1"
} | null | null | new_dataset | admin | null | false | null | 5f81a480-6c41-422d-a3f5-cf0ea5d181b3 | null | Validated | {
"text_length": 1471
} | 0new_dataset
|
TITLE: Semi-Supervised Semantic Segmentation in Earth Observation: The MiniFrance Suite, Dataset Analysis and Multi-task Network Study
ABSTRACT: The development of semi-supervised learning techniques is essential to
enhance the generalization capacities of machine learning algorithms. Indeed,
raw image data are abundant while labels are scarce, therefore it is crucial to
leverage unlabeled inputs to build better models. The availability of large
databases have been key for the development of learning algorithms with high
level performance.
Despite the major role of machine learning in Earth Observation to derive
products such as land cover maps, datasets in the field are still limited,
either because of modest surface coverage, lack of variety of scenes or
restricted classes to identify. We introduce a novel large-scale dataset for
semi-supervised semantic segmentation in Earth Observation, the MiniFrance
suite. MiniFrance has several unprecedented properties: it is large-scale,
containing over 2000 very high resolution aerial images, accounting for more
than 200 billions samples (pixels); it is varied, covering 16 conurbations in
France, with various climates, different landscapes, and urban as well as
countryside scenes; and it is challenging, considering land use classes with
high-level semantics. Nevertheless, the most distinctive quality of MiniFrance
is being the only dataset in the field especially designed for semi-supervised
learning: it contains labeled and unlabeled images in its training partition,
which reproduces a life-like scenario. Along with this dataset, we present
tools for data representativeness analysis in terms of appearance similarity
and a thorough study of MiniFrance data, demonstrating that it is suitable for
learning and generalizes well in a semi-supervised setting. Finally, we present
semi-supervised deep architectures based on multi-task learning and the first
experiments on MiniFrance. | {
"abstract": "The development of semi-supervised learning techniques is essential to\nenhance the generalization capacities of machine learning algorithms. Indeed,\nraw image data are abundant while labels are scarce, therefore it is crucial to\nleverage unlabeled inputs to build better models. The availability of large\ndatabases have been key for the development of learning algorithms with high\nlevel performance.\n Despite the major role of machine learning in Earth Observation to derive\nproducts such as land cover maps, datasets in the field are still limited,\neither because of modest surface coverage, lack of variety of scenes or\nrestricted classes to identify. We introduce a novel large-scale dataset for\nsemi-supervised semantic segmentation in Earth Observation, the MiniFrance\nsuite. MiniFrance has several unprecedented properties: it is large-scale,\ncontaining over 2000 very high resolution aerial images, accounting for more\nthan 200 billions samples (pixels); it is varied, covering 16 conurbations in\nFrance, with various climates, different landscapes, and urban as well as\ncountryside scenes; and it is challenging, considering land use classes with\nhigh-level semantics. Nevertheless, the most distinctive quality of MiniFrance\nis being the only dataset in the field especially designed for semi-supervised\nlearning: it contains labeled and unlabeled images in its training partition,\nwhich reproduces a life-like scenario. Along with this dataset, we present\ntools for data representativeness analysis in terms of appearance similarity\nand a thorough study of MiniFrance data, demonstrating that it is suitable for\nlearning and generalizes well in a semi-supervised setting. Finally, we present\nsemi-supervised deep architectures based on multi-task learning and the first\nexperiments on MiniFrance.",
"title": "Semi-Supervised Semantic Segmentation in Earth Observation: The MiniFrance Suite, Dataset Analysis and Multi-task Network Study",
"url": "http://arxiv.org/abs/2010.07830v1"
} | null | null | new_dataset | admin | null | false | null | 27070983-724c-4a1f-b90c-e1d8738d2816 | null | Validated | {
"text_length": 1970
} | 0new_dataset
|
TITLE: RaidaR: A Rich Annotated Image Dataset of Rainy Street Scenes
ABSTRACT: We introduce RaidaR, a rich annotated image dataset of rainy street scenes,
to support autonomous driving research. The new dataset contains the largest
number of rainy images (58,542) to date, 5,000 of which provide semantic
segmentations and 3,658 provide object instance segmentations. The RaidaR
images cover a wide range of realistic rain-induced artifacts, including fog,
droplets, and road reflections, which can effectively augment existing street
scene datasets to improve data-driven machine perception during rainy weather.
To facilitate efficient annotation of a large volume of images, we develop a
semi-automatic scheme combining manual segmentation and an automated processing
akin to cross validation, resulting in 10-20 fold reduction on annotation time.
We demonstrate the utility of our new dataset by showing how data augmentation
with RaidaR can elevate the accuracy of existing segmentation algorithms. We
also present a novel unpaired image-to-image translation algorithm for
adding/removing rain artifacts, which directly benefits from RaidaR. | {
"abstract": "We introduce RaidaR, a rich annotated image dataset of rainy street scenes,\nto support autonomous driving research. The new dataset contains the largest\nnumber of rainy images (58,542) to date, 5,000 of which provide semantic\nsegmentations and 3,658 provide object instance segmentations. The RaidaR\nimages cover a wide range of realistic rain-induced artifacts, including fog,\ndroplets, and road reflections, which can effectively augment existing street\nscene datasets to improve data-driven machine perception during rainy weather.\nTo facilitate efficient annotation of a large volume of images, we develop a\nsemi-automatic scheme combining manual segmentation and an automated processing\nakin to cross validation, resulting in 10-20 fold reduction on annotation time.\nWe demonstrate the utility of our new dataset by showing how data augmentation\nwith RaidaR can elevate the accuracy of existing segmentation algorithms. We\nalso present a novel unpaired image-to-image translation algorithm for\nadding/removing rain artifacts, which directly benefits from RaidaR.",
"title": "RaidaR: A Rich Annotated Image Dataset of Rainy Street Scenes",
"url": "http://arxiv.org/abs/2104.04606v3"
} | null | null | new_dataset | admin | null | false | null | 2811a11c-72ae-43e1-bf62-d086501ece10 | null | Validated | {
"text_length": 1163
} | 0new_dataset
|
TITLE: AIR-Act2Act: Human-human interaction dataset for teaching non-verbal social behaviors to robots
ABSTRACT: To better interact with users, a social robot should understand the users'
behavior, infer the intention, and respond appropriately. Machine learning is
one way of implementing robot intelligence. It provides the ability to
automatically learn and improve from experience instead of explicitly telling
the robot what to do. Social skills can also be learned through watching
human-human interaction videos. However, human-human interaction datasets are
relatively scarce to learn interactions that occur in various situations.
Moreover, we aim to use service robots in the elderly-care domain; however,
there has been no interaction dataset collected for this domain. For this
reason, we introduce a human-human interaction dataset for teaching non-verbal
social behaviors to robots. It is the only interaction dataset that elderly
people have participated in as performers. We recruited 100 elderly people and
two college students to perform 10 interactions in an indoor environment. The
entire dataset has 5,000 interaction samples, each of which contains depth
maps, body indexes and 3D skeletal data that are captured with three Microsoft
Kinect v2 cameras. In addition, we provide the joint angles of a humanoid NAO
robot which are converted from the human behavior that robots need to learn.
The dataset and useful python scripts are available for download at
https://github.com/ai4r/AIR-Act2Act. It can be used to not only teach social
skills to robots but also benchmark action recognition algorithms. | {
"abstract": "To better interact with users, a social robot should understand the users'\nbehavior, infer the intention, and respond appropriately. Machine learning is\none way of implementing robot intelligence. It provides the ability to\nautomatically learn and improve from experience instead of explicitly telling\nthe robot what to do. Social skills can also be learned through watching\nhuman-human interaction videos. However, human-human interaction datasets are\nrelatively scarce to learn interactions that occur in various situations.\nMoreover, we aim to use service robots in the elderly-care domain; however,\nthere has been no interaction dataset collected for this domain. For this\nreason, we introduce a human-human interaction dataset for teaching non-verbal\nsocial behaviors to robots. It is the only interaction dataset that elderly\npeople have participated in as performers. We recruited 100 elderly people and\ntwo college students to perform 10 interactions in an indoor environment. The\nentire dataset has 5,000 interaction samples, each of which contains depth\nmaps, body indexes and 3D skeletal data that are captured with three Microsoft\nKinect v2 cameras. In addition, we provide the joint angles of a humanoid NAO\nrobot which are converted from the human behavior that robots need to learn.\nThe dataset and useful python scripts are available for download at\nhttps://github.com/ai4r/AIR-Act2Act. It can be used to not only teach social\nskills to robots but also benchmark action recognition algorithms.",
"title": "AIR-Act2Act: Human-human interaction dataset for teaching non-verbal social behaviors to robots",
"url": "http://arxiv.org/abs/2009.02041v1"
} | null | null | new_dataset | admin | null | false | null | 0aa0a75f-ff3b-4620-9932-64ad6aea89e4 | null | Validated | {
"text_length": 1639
} | 0new_dataset
|
TITLE: Interpreting Black-box Machine Learning Models for High Dimensional Datasets
ABSTRACT: Deep neural networks (DNNs) have been shown to outperform traditional machine
learning algorithms in a broad variety of application domains due to their
effectiveness in modeling complex problems and handling high-dimensional
datasets. Many real-life datasets, however, are of increasingly high
dimensionality, where a large number of features may be irrelevant for both
supervised and unsupervised learning tasks. The inclusion of such features
would not only introduce unwanted noise but also increase computational
complexity. Furthermore, due to high non-linearity and dependency among a large
number of features, DNN models tend to be unavoidably opaque and perceived as
black-box methods because of their not well-understood internal functioning.
Their algorithmic complexity is often simply beyond the capacities of humans to
understand the interplay among myriads of hyperparameters. A well-interpretable
model can identify statistically significant features and explain the way they
affect the model's outcome. In this paper, we propose an efficient method to
improve the interpretability of black-box models for classification tasks in
the case of high-dimensional datasets. First, we train a black-box model on a
high-dimensional dataset to learn the embeddings on which the classification is
performed. To decompose the inner working principles of the black-box model and
to identify top-k important features, we employ different probing and
perturbing techniques. We then approximate the behavior of the black-box model
by means of an interpretable surrogate model on the top-k feature space.
Finally, we derive decision rules and local explanations from the surrogate
model to explain individual decisions. Our approach outperforms
state-of-the-art methods like TabNet and XGboost when tested on different
datasets with varying dimensionality between 50 and 20,000 w.r.t metrics and
explainability. | {
"abstract": "Deep neural networks (DNNs) have been shown to outperform traditional machine\nlearning algorithms in a broad variety of application domains due to their\neffectiveness in modeling complex problems and handling high-dimensional\ndatasets. Many real-life datasets, however, are of increasingly high\ndimensionality, where a large number of features may be irrelevant for both\nsupervised and unsupervised learning tasks. The inclusion of such features\nwould not only introduce unwanted noise but also increase computational\ncomplexity. Furthermore, due to high non-linearity and dependency among a large\nnumber of features, DNN models tend to be unavoidably opaque and perceived as\nblack-box methods because of their not well-understood internal functioning.\nTheir algorithmic complexity is often simply beyond the capacities of humans to\nunderstand the interplay among myriads of hyperparameters. A well-interpretable\nmodel can identify statistically significant features and explain the way they\naffect the model's outcome. In this paper, we propose an efficient method to\nimprove the interpretability of black-box models for classification tasks in\nthe case of high-dimensional datasets. First, we train a black-box model on a\nhigh-dimensional dataset to learn the embeddings on which the classification is\nperformed. To decompose the inner working principles of the black-box model and\nto identify top-k important features, we employ different probing and\nperturbing techniques. We then approximate the behavior of the black-box model\nby means of an interpretable surrogate model on the top-k feature space.\nFinally, we derive decision rules and local explanations from the surrogate\nmodel to explain individual decisions. Our approach outperforms\nstate-of-the-art methods like TabNet and XGboost when tested on different\ndatasets with varying dimensionality between 50 and 20,000 w.r.t metrics and\nexplainability.",
"title": "Interpreting Black-box Machine Learning Models for High Dimensional Datasets",
"url": "http://arxiv.org/abs/2208.13405v2"
} | null | null | no_new_dataset | admin | null | false | null | a7139c48-a50b-4a25-b653-75cfe55224fe | null | Validated | {
"text_length": 2023
} | 1no_new_dataset
|
TITLE: XLCoST: A Benchmark Dataset for Cross-lingual Code Intelligence
ABSTRACT: Recent advances in machine learning have significantly improved the
understanding of source code data and achieved good performance on a number of
downstream tasks. Open source repositories like GitHub enable this process with
rich unlabeled code data. However, the lack of high quality labeled data has
largely hindered the progress of several code related tasks, such as program
translation, summarization, synthesis, and code search. This paper introduces
XLCoST, Cross-Lingual Code SnippeT dataset, a new benchmark dataset for
cross-lingual code intelligence. Our dataset contains fine-grained parallel
data from 8 languages (7 commonly used programming languages and English), and
supports 10 cross-lingual code tasks. To the best of our knowledge, it is the
largest parallel dataset for source code both in terms of size and the number
of languages. We also provide the performance of several state-of-the-art
baseline models for each task. We believe this new dataset can be a valuable
asset for the research community and facilitate the development and validation
of new methods for cross-lingual code intelligence. | {
"abstract": "Recent advances in machine learning have significantly improved the\nunderstanding of source code data and achieved good performance on a number of\ndownstream tasks. Open source repositories like GitHub enable this process with\nrich unlabeled code data. However, the lack of high quality labeled data has\nlargely hindered the progress of several code related tasks, such as program\ntranslation, summarization, synthesis, and code search. This paper introduces\nXLCoST, Cross-Lingual Code SnippeT dataset, a new benchmark dataset for\ncross-lingual code intelligence. Our dataset contains fine-grained parallel\ndata from 8 languages (7 commonly used programming languages and English), and\nsupports 10 cross-lingual code tasks. To the best of our knowledge, it is the\nlargest parallel dataset for source code both in terms of size and the number\nof languages. We also provide the performance of several state-of-the-art\nbaseline models for each task. We believe this new dataset can be a valuable\nasset for the research community and facilitate the development and validation\nof new methods for cross-lingual code intelligence.",
"title": "XLCoST: A Benchmark Dataset for Cross-lingual Code Intelligence",
"url": "http://arxiv.org/abs/2206.08474v1"
} | null | null | new_dataset | admin | null | false | null | a105b929-87c4-4981-99c6-4d6687c61b36 | null | Validated | {
"text_length": 1221
} | 0new_dataset
|
TITLE: IFCNet: A Benchmark Dataset for IFC Entity Classification
ABSTRACT: Enhancing interoperability and information exchange between domain-specific
software products for BIM is an important aspect in the Architecture,
Engineering, Construction and Operations industry. Recent research started
investigating methods from the areas of machine and deep learning for semantic
enrichment of BIM models. However, training and evaluation of these machine
learning algorithms requires sufficiently large and comprehensive datasets.
This work presents IFCNet, a dataset of single-entity IFC files spanning a
broad range of IFC classes containing both geometric and semantic information.
Using only the geometric information of objects, the experiments show that
three different deep learning models are able to achieve good classification
performance. | {
"abstract": "Enhancing interoperability and information exchange between domain-specific\nsoftware products for BIM is an important aspect in the Architecture,\nEngineering, Construction and Operations industry. Recent research started\ninvestigating methods from the areas of machine and deep learning for semantic\nenrichment of BIM models. However, training and evaluation of these machine\nlearning algorithms requires sufficiently large and comprehensive datasets.\nThis work presents IFCNet, a dataset of single-entity IFC files spanning a\nbroad range of IFC classes containing both geometric and semantic information.\nUsing only the geometric information of objects, the experiments show that\nthree different deep learning models are able to achieve good classification\nperformance.",
"title": "IFCNet: A Benchmark Dataset for IFC Entity Classification",
"url": "http://arxiv.org/abs/2106.09712v1"
} | null | null | new_dataset | admin | null | false | null | e6fb60d6-17c4-4d32-a8b2-8666742d1ee0 | null | Validated | {
"text_length": 862
} | 0new_dataset
|
TITLE: Intrusion Detection Systems Using Support Vector Machines on the KDDCUP'99 and NSL-KDD Datasets: A Comprehensive Survey
ABSTRACT: With the growing rates of cyber-attacks and cyber espionage, the need for
better and more powerful intrusion detection systems (IDS) is even more
warranted nowadays. The basic task of an IDS is to act as the first line of
defense, in detecting attacks on the internet. As intrusion tactics from
intruders become more sophisticated and difficult to detect, researchers have
started to apply novel Machine Learning (ML) techniques to effectively detect
intruders and hence preserve internet users' information and overall trust in
the entire internet network security. Over the last decade, there has been an
explosion of research on intrusion detection techniques based on ML and Deep
Learning (DL) architectures on various cyber security-based datasets such as
the DARPA, KDDCUP'99, NSL-KDD, CAIDA, CTU-13, UNSW-NB15. In this research, we
review contemporary literature and provide a comprehensive survey of different
types of intrusion detection technique that applies Support Vector Machines
(SVMs) algorithms as a classifier. We focus only on studies that have been
evaluated on the two most widely used datasets in cybersecurity namely: the
KDDCUP'99 and the NSL-KDD datasets. We provide a summary of each method,
identifying the role of the SVMs classifier, and all other algorithms involved
in the studies. Furthermore, we present a critical review of each method, in
tabular form, highlighting the performance measures, strengths, and limitations
of each of the methods surveyed. | {
"abstract": "With the growing rates of cyber-attacks and cyber espionage, the need for\nbetter and more powerful intrusion detection systems (IDS) is even more\nwarranted nowadays. The basic task of an IDS is to act as the first line of\ndefense, in detecting attacks on the internet. As intrusion tactics from\nintruders become more sophisticated and difficult to detect, researchers have\nstarted to apply novel Machine Learning (ML) techniques to effectively detect\nintruders and hence preserve internet users' information and overall trust in\nthe entire internet network security. Over the last decade, there has been an\nexplosion of research on intrusion detection techniques based on ML and Deep\nLearning (DL) architectures on various cyber security-based datasets such as\nthe DARPA, KDDCUP'99, NSL-KDD, CAIDA, CTU-13, UNSW-NB15. In this research, we\nreview contemporary literature and provide a comprehensive survey of different\ntypes of intrusion detection technique that applies Support Vector Machines\n(SVMs) algorithms as a classifier. We focus only on studies that have been\nevaluated on the two most widely used datasets in cybersecurity namely: the\nKDDCUP'99 and the NSL-KDD datasets. We provide a summary of each method,\nidentifying the role of the SVMs classifier, and all other algorithms involved\nin the studies. Furthermore, we present a critical review of each method, in\ntabular form, highlighting the performance measures, strengths, and limitations\nof each of the methods surveyed.",
"title": "Intrusion Detection Systems Using Support Vector Machines on the KDDCUP'99 and NSL-KDD Datasets: A Comprehensive Survey",
"url": "http://arxiv.org/abs/2209.05579v1"
} | null | null | no_new_dataset | admin | null | false | null | 0b5e25e8-a4cd-4de2-818b-046552e7461f | null | Validated | {
"text_length": 1640
} | 1no_new_dataset
|
TITLE: NEREL-BIO: A Dataset of Biomedical Abstracts Annotated with Nested Named Entities
ABSTRACT: This paper describes NEREL-BIO -- an annotation scheme and corpus of PubMed
abstracts in Russian and smaller number of abstracts in English. NEREL-BIO
extends the general domain dataset NEREL by introducing domain-specific entity
types. NEREL-BIO annotation scheme covers both general and biomedical domains
making it suitable for domain transfer experiments. NEREL-BIO provides
annotation for nested named entities as an extension of the scheme employed for
NEREL. Nested named entities may cross entity boundaries to connect to shorter
entities nested within longer entities, making them harder to detect.
NEREL-BIO contains annotations for 700+ Russian and 100+ English abstracts.
All English PubMed annotations have corresponding Russian counterparts. Thus,
NEREL-BIO comprises the following specific features: annotation of nested named
entities, it can be used as a benchmark for cross-domain (NEREL -> NEREL-BIO)
and cross-language (English -> Russian) transfer. We experiment with both
transformer-based sequence models and machine reading comprehension (MRC)
models and report their results.
The dataset is freely available at https://github.com/nerel-ds/NEREL-BIO. | {
"abstract": "This paper describes NEREL-BIO -- an annotation scheme and corpus of PubMed\nabstracts in Russian and smaller number of abstracts in English. NEREL-BIO\nextends the general domain dataset NEREL by introducing domain-specific entity\ntypes. NEREL-BIO annotation scheme covers both general and biomedical domains\nmaking it suitable for domain transfer experiments. NEREL-BIO provides\nannotation for nested named entities as an extension of the scheme employed for\nNEREL. Nested named entities may cross entity boundaries to connect to shorter\nentities nested within longer entities, making them harder to detect.\n NEREL-BIO contains annotations for 700+ Russian and 100+ English abstracts.\nAll English PubMed annotations have corresponding Russian counterparts. Thus,\nNEREL-BIO comprises the following specific features: annotation of nested named\nentities, it can be used as a benchmark for cross-domain (NEREL -> NEREL-BIO)\nand cross-language (English -> Russian) transfer. We experiment with both\ntransformer-based sequence models and machine reading comprehension (MRC)\nmodels and report their results.\n The dataset is freely available at https://github.com/nerel-ds/NEREL-BIO.",
"title": "NEREL-BIO: A Dataset of Biomedical Abstracts Annotated with Nested Named Entities",
"url": "http://arxiv.org/abs/2210.11913v1"
} | null | null | new_dataset | admin | null | false | null | 17db3f79-0176-40de-8448-c76da8578952 | null | Validated | {
"text_length": 1294
} | 0new_dataset
|
TITLE: Towards emotion recognition for virtual environments: an evaluation of EEG features on benchmark dataset
ABSTRACT: One of the challenges in virtual environments is the difficulty users have in
interacting with these increasingly complex systems. Ultimately, endowing
machines with the ability to perceive users emotions will enable a more
intuitive and reliable interaction. Consequently, using the
electroencephalogram as a bio-signal sensor, the affective state of a user can
be modelled and subsequently utilised in order to achieve a system that can
recognise and react to the user's emotions. This paper investigates features
extracted from electroencephalogram signals for the purpose of affective state
modelling based on Russell's Circumplex Model. Investigations are presented
that aim to provide the foundation for future work in modelling user affect to
enhance interaction experience in virtual environments. The DEAP dataset was
used within this work, along with a Support Vector Machine and Random Forest,
which yielded reasonable classification accuracies for Valence and Arousal
using feature vectors based on statistical measurements and band power from the
\'z, \b{eta}, \'z, and \'z\'z waves and High Order Crossing of the EEG signal. | {
"abstract": "One of the challenges in virtual environments is the difficulty users have in\ninteracting with these increasingly complex systems. Ultimately, endowing\nmachines with the ability to perceive users emotions will enable a more\nintuitive and reliable interaction. Consequently, using the\nelectroencephalogram as a bio-signal sensor, the affective state of a user can\nbe modelled and subsequently utilised in order to achieve a system that can\nrecognise and react to the user's emotions. This paper investigates features\nextracted from electroencephalogram signals for the purpose of affective state\nmodelling based on Russell's Circumplex Model. Investigations are presented\nthat aim to provide the foundation for future work in modelling user affect to\nenhance interaction experience in virtual environments. The DEAP dataset was\nused within this work, along with a Support Vector Machine and Random Forest,\nwhich yielded reasonable classification accuracies for Valence and Arousal\nusing feature vectors based on statistical measurements and band power from the\n\\'z, \\b{eta}, \\'z, and \\'z\\'z waves and High Order Crossing of the EEG signal.",
"title": "Towards emotion recognition for virtual environments: an evaluation of EEG features on benchmark dataset",
"url": "http://arxiv.org/abs/2210.13876v1"
} | null | null | no_new_dataset | admin | null | false | null | e755446b-2f17-4138-8928-166ca1d0359a | null | Validated | {
"text_length": 1277
} | 1no_new_dataset
|
TITLE: METS-CoV: A Dataset of Medical Entity and Targeted Sentiment on COVID-19 Related Tweets
ABSTRACT: The COVID-19 pandemic continues to bring up various topics discussed or
debated on social media. In order to explore the impact of pandemics on
people's lives, it is crucial to understand the public's concerns and attitudes
towards pandemic-related entities (e.g., drugs, vaccines) on social media.
However, models trained on existing named entity recognition (NER) or targeted
sentiment analysis (TSA) datasets have limited ability to understand
COVID-19-related social media texts because these datasets are not designed or
annotated from a medical perspective. This paper releases METS-CoV, a dataset
containing medical entities and targeted sentiments from COVID-19-related
tweets. METS-CoV contains 10,000 tweets with 7 types of entities, including 4
medical entity types (Disease, Drug, Symptom, and Vaccine) and 3 general entity
types (Person, Location, and Organization). To further investigate tweet users'
attitudes toward specific entities, 4 types of entities (Person, Organization,
Drug, and Vaccine) are selected and annotated with user sentiments, resulting
in a targeted sentiment dataset with 9,101 entities (in 5,278 tweets). To the
best of our knowledge, METS-CoV is the first dataset to collect medical
entities and corresponding sentiments of COVID-19-related tweets. We benchmark
the performance of classical machine learning models and state-of-the-art deep
learning models on NER and TSA tasks with extensive experiments. Results show
that the dataset has vast room for improvement for both NER and TSA tasks.
METS-CoV is an important resource for developing better medical social media
tools and facilitating computational social science research, especially in
epidemiology. Our data, annotation guidelines, benchmark models, and source
code are publicly available (https://github.com/YLab-Open/METS-CoV) to ensure
reproducibility. | {
"abstract": "The COVID-19 pandemic continues to bring up various topics discussed or\ndebated on social media. In order to explore the impact of pandemics on\npeople's lives, it is crucial to understand the public's concerns and attitudes\ntowards pandemic-related entities (e.g., drugs, vaccines) on social media.\nHowever, models trained on existing named entity recognition (NER) or targeted\nsentiment analysis (TSA) datasets have limited ability to understand\nCOVID-19-related social media texts because these datasets are not designed or\nannotated from a medical perspective. This paper releases METS-CoV, a dataset\ncontaining medical entities and targeted sentiments from COVID-19-related\ntweets. METS-CoV contains 10,000 tweets with 7 types of entities, including 4\nmedical entity types (Disease, Drug, Symptom, and Vaccine) and 3 general entity\ntypes (Person, Location, and Organization). To further investigate tweet users'\nattitudes toward specific entities, 4 types of entities (Person, Organization,\nDrug, and Vaccine) are selected and annotated with user sentiments, resulting\nin a targeted sentiment dataset with 9,101 entities (in 5,278 tweets). To the\nbest of our knowledge, METS-CoV is the first dataset to collect medical\nentities and corresponding sentiments of COVID-19-related tweets. We benchmark\nthe performance of classical machine learning models and state-of-the-art deep\nlearning models on NER and TSA tasks with extensive experiments. Results show\nthat the dataset has vast room for improvement for both NER and TSA tasks.\nMETS-CoV is an important resource for developing better medical social media\ntools and facilitating computational social science research, especially in\nepidemiology. Our data, annotation guidelines, benchmark models, and source\ncode are publicly available (https://github.com/YLab-Open/METS-CoV) to ensure\nreproducibility.",
"title": "METS-CoV: A Dataset of Medical Entity and Targeted Sentiment on COVID-19 Related Tweets",
"url": "http://arxiv.org/abs/2209.13773v1"
} | null | null | new_dataset | admin | null | false | null | 401077d3-09e9-4b72-914e-7e0d65b3e45d | null | Validated | {
"text_length": 1979
} | 0new_dataset
|
TITLE: A Benchmarking Dataset with 2440 Organic Molecules for Volume Distribution at Steady State
ABSTRACT: Background: The volume of distribution at steady state (VDss) is a
fundamental pharmacokinetics (PK) property of drugs, which measures how
effectively a drug molecule is distributed throughout the body. Along with the
clearance (CL), it determines the half-life and, therefore, the drug dosing
interval. However, the molecular data size limits the generalizability of the
reported machine learning models. Objective: This study aims to provide a clean
and comprehensive dataset for human VDss as the benchmarking data source,
fostering and benefiting future predictive studies. Moreover, several
predictive models were also built with machine learning regression algorithms.
Methods: The dataset was curated from 13 publicly accessible data sources and
the DrugBank database entirely from intravenous drug administration and then
underwent extensive data cleaning. The molecular descriptors were calculated
with Mordred, and feature selection was conducted for constructing predictive
models. Five machine learning methods were used to build regression models,
grid search was used to optimize hyperparameters, and ten-fold cross-validation
was used to evaluate the model. Results: An enriched dataset of VDss
(https://github.com/da-wen-er/VDss) was constructed with 2440 molecules. Among
the prediction models, the LightGBM model was the most stable and had the best
internal prediction ability with Q2 = 0.837, R2=0.814 and for the other four
models, Q2 was higher than 0.79. Conclusions: To the best of our knowledge,
this is the largest dataset for VDss, which can be used as the benchmark for
computational studies of VDss. Moreover, the regression models reported within
this study can be of use for pharmacokinetic related studies. | {
"abstract": "Background: The volume of distribution at steady state (VDss) is a\nfundamental pharmacokinetics (PK) property of drugs, which measures how\neffectively a drug molecule is distributed throughout the body. Along with the\nclearance (CL), it determines the half-life and, therefore, the drug dosing\ninterval. However, the molecular data size limits the generalizability of the\nreported machine learning models. Objective: This study aims to provide a clean\nand comprehensive dataset for human VDss as the benchmarking data source,\nfostering and benefiting future predictive studies. Moreover, several\npredictive models were also built with machine learning regression algorithms.\nMethods: The dataset was curated from 13 publicly accessible data sources and\nthe DrugBank database entirely from intravenous drug administration and then\nunderwent extensive data cleaning. The molecular descriptors were calculated\nwith Mordred, and feature selection was conducted for constructing predictive\nmodels. Five machine learning methods were used to build regression models,\ngrid search was used to optimize hyperparameters, and ten-fold cross-validation\nwas used to evaluate the model. Results: An enriched dataset of VDss\n(https://github.com/da-wen-er/VDss) was constructed with 2440 molecules. Among\nthe prediction models, the LightGBM model was the most stable and had the best\ninternal prediction ability with Q2 = 0.837, R2=0.814 and for the other four\nmodels, Q2 was higher than 0.79. Conclusions: To the best of our knowledge,\nthis is the largest dataset for VDss, which can be used as the benchmark for\ncomputational studies of VDss. Moreover, the regression models reported within\nthis study can be of use for pharmacokinetic related studies.",
"title": "A Benchmarking Dataset with 2440 Organic Molecules for Volume Distribution at Steady State",
"url": "http://arxiv.org/abs/2211.05661v1"
} | null | null | new_dataset | admin | null | false | null | 4be8cf84-9ee8-452c-8270-486868b8c99f | null | Validated | {
"text_length": 1863
} | 0new_dataset
|
TITLE: SyntheticFur dataset for neural rendering
ABSTRACT: We introduce a new dataset called SyntheticFur built specifically for machine
learning training. The dataset consists of ray traced synthetic fur renders
with corresponding rasterized input buffers and simulation data files. We
procedurally generated approximately 140,000 images and 15 simulations with
Houdini. The images consist of fur groomed with different skin primitives and
move with various motions in a predefined set of lighting environments. We also
demonstrated how the dataset could be used with neural rendering to
significantly improve fur graphics using inexpensive input buffers by training
a conditional generative adversarial network with perceptual loss. We hope the
availability of such high fidelity fur renders will encourage new advances with
neural rendering for a variety of applications. | {
"abstract": "We introduce a new dataset called SyntheticFur built specifically for machine\nlearning training. The dataset consists of ray traced synthetic fur renders\nwith corresponding rasterized input buffers and simulation data files. We\nprocedurally generated approximately 140,000 images and 15 simulations with\nHoudini. The images consist of fur groomed with different skin primitives and\nmove with various motions in a predefined set of lighting environments. We also\ndemonstrated how the dataset could be used with neural rendering to\nsignificantly improve fur graphics using inexpensive input buffers by training\na conditional generative adversarial network with perceptual loss. We hope the\navailability of such high fidelity fur renders will encourage new advances with\nneural rendering for a variety of applications.",
"title": "SyntheticFur dataset for neural rendering",
"url": "http://arxiv.org/abs/2105.06409v1"
} | null | null | new_dataset | admin | null | false | null | 1eb85054-81ea-44c4-ad59-bfe6c5ac9d31 | null | Validated | {
"text_length": 891
} | 0new_dataset
|
TITLE: Seeing the Unseen: Errors and Bias in Visual Datasets
ABSTRACT: From face recognition in smartphones to automatic routing on self-driving
cars, machine vision algorithms lie in the core of these features. These
systems solve image based tasks by identifying and understanding objects,
subsequently making decisions from these information. However, errors in
datasets are usually induced or even magnified in algorithms, at times
resulting in issues such as recognising black people as gorillas and
misrepresenting ethnicities in search results. This paper tracks the errors in
datasets and their impacts, revealing that a flawed dataset could be a result
of limited categories, incomprehensive sourcing and poor classification. | {
"abstract": "From face recognition in smartphones to automatic routing on self-driving\ncars, machine vision algorithms lie in the core of these features. These\nsystems solve image based tasks by identifying and understanding objects,\nsubsequently making decisions from these information. However, errors in\ndatasets are usually induced or even magnified in algorithms, at times\nresulting in issues such as recognising black people as gorillas and\nmisrepresenting ethnicities in search results. This paper tracks the errors in\ndatasets and their impacts, revealing that a flawed dataset could be a result\nof limited categories, incomprehensive sourcing and poor classification.",
"title": "Seeing the Unseen: Errors and Bias in Visual Datasets",
"url": "http://arxiv.org/abs/2211.01847v1"
} | null | null | no_new_dataset | admin | null | false | null | 122d9848-2c69-4636-994b-dc7dcf9f7f5f | null | Validated | {
"text_length": 751
} | 1no_new_dataset
|
End of preview. Expand
in Dataset Viewer.
No dataset card yet
New: Create and edit this dataset card directly on the website!
Contribute a Dataset Card- Downloads last month
- 10