Solutions recap

#8
by Ilde - opened

Hello! Would be cool if you guys @VincentSchuler and @fconti commented on your approach to the problem :D. I'm somewhat new to competitions and just would like to know how you went about it, just to learn a bit. As for me, I ended up using an LLM to concatenate the used rules (the text, not the numeric rules) by users and then predicting possible labels (text as well). I figured the rules made some semantic sense and that people ended up using similar rules most of the time with the devices they had already used. The end was a bit messy as I had to convert this text rule to the numeric one by deciding which devices to use. It worked fairly well, though I did manage to squeeze a much better score.

Anyways, had fun with it. Hope to hear from you guys. Cheers

Hello @Ilde !

Good game!

your solution seems very different from mine and also from Vincent's (which I reached on linkedin). it's nice that the top 3 have totally different approaches.

I am relatively new to the world of competitions, I developed a GNN based on the baseline paper published by wyze (https://openreview.net/forum?id=qynH28Y4xE).

Basically, I modeled the problem as a link prediction, in which a score is assigned to each possible link of each type. The architecture consists of a first embedding block for each node of the graph and then a head for calculating the probabilities for each of the links. So I selected the 50 best scores. I trained using positive and negative sampling from the training set.

The solution is quite similar to that of wyze although I re-implemented everything using pytorch geometric. I think the biggest differences:

  • Training organized in batches;
  • Positive and negative sampling different from the baseline;
  • The mechanism for defining features on nodes and links.

In any case, I haven't had time to do hyper-parameter tuning, so I think the performance can improve a lot. I worked just with free google-colab.
I will definitely share the solution on github!

Oh cool, looking forward to it. Thanks for the write up. I think I used that just to toy with the sample submission, but was really looking for an excuse to learn more about the transformers library. Cheers

Oh cool! I'll check it out. Cheers

Sign up or log in to comment