TOPOAUC

An implement of the ACM MM 22 paper: A Unified Framework against Topology and Class Imbalance.

Environments

• Python 3.7.11
• Pytorch 1.11
• torch-geometric 2.0.4
• CUDA 11.3
• scikit-learn 1.0.2

Data

When running train.py for the first time, the dataset (CORA, CiteSeer, PubMed) will be automatically downloaded to ./datasets/[dataset] by torch_geometric.

Training

1. Modify the config file config.py (copy the parameters from ./best_params/layers3/[dataset]/search_space_imb_losses_[loss type]_[class imbalance ratio].json)
2. Run the script:

CUDA_VISIBLE_DEVICES=0  python train.py --loss ExpGAUC --pair_ner_diff 1 --imb_ratio 10.0

Note: ["ExpGAUC","HingeGAUC","SqGAUC"] are three kinds of AUC losses, --pair_ner_diff decides whether to use our TAIL mechanism, --imb_ratio controls the class imbalance ratio which could be selected from [10.0,15.0,20.0].