In this repo, we provide our solution to KDD cup 2021. We use three types of models:

• GIN network, which is the GIN with virtual nodes. The code is in the folder GIN;
• Standard Transformer, which is the same as that in NLP. The code is in the folder Standard Transformer;
• Two-branch Transformer, which is a variant of Transformer with a regression branch and a classification branch. The code is in the folder Two-branch Transformer.

Please refer to the README in each folder for more details.

Our pre-trained models are released on Onedrive.

Our all prediction are released at final_v0.

If you want to reproduce our results, go each subdirectory and follow the instructions to reconstruct the results. After that, by running submitted_predictions.py, you can merge all predictions of each type of model.

The data processing steps are shown in DataPreprocessing folder.

For technique details, see this paper.