Rules are simple: No rules. You start with two identical tasks from the Open Graph Benchmark from Stanford. Two tasks are selected: Node classification with ogbn-arxiv and graph classification with ogbg-molhiv.

If first in one task, you'll earn a mystery legendary loot. Hence, two winners (or one) will be elected. Your final grade will simply be related to your participation to this challenge. Submission to this benchmark are simple as well, upload directory with your name on the corresponding folder e.g. arxiv/parcollet-titouan. Then, add your score to the corresponding table in the Readme.md!

YOUR CODE MUST BE SELF-CONTAINED I am fine with installing dependencies but basically, I do not expect to see more than 4 files in your submission: requirements.txt (for external dependencies), train.py (the trainer), model.py (definition of the model) and data.py (for any RELEVANT data preparation).

any entry to the leaderboard that isn't reproducible will be invalidated by myself (non-reproducible research sucks)!

• Start small, explore the dataset, understand it.
• Still start small: implement a simple GCN BY YOURSELF. This will make you confident in moving forward.
• Do not blindly apply all the models from torch_geometric. Tuning a promising model instead of trying thousands of them will certainly lead you to the top-1. Plus, I won't be happy if your code just contain a single GNN from PyTorch-Geometric. I could be happy if you try 10 of then and benchmark them properly (statistics etc ...).
• Do not think that a custom model is a bad idea. Models implemented by default in torch_geometric are generic. It is most likely that a SotA handcrafted model will end-up being better !