Rethinking Graph Neural Networks for Anomaly Detection

This is the official implementation for the following paper:

Rethinking Graph Neural Networks for Anomaly Detection
Jianheng Tang, Jiajin Li, Ziqi Gao, Jia Li
ICML 2022

BWGNN has been integrated into GADBench, a comprehensive benchmark for (semi-)supervised graph anomaly detection.

Dependencies

• pytorch 1.9.0
• dgl 0.8.1
• sympy
• argparse
• sklearn

How to run

The T-Finance and T-Social datasets developed in the paper are on google drive. Download and unzip all files in the dataset folder.

plot.zip in the above link is used to reproduce Figure 1 and 2 in our paper. You can unzip it and directly run the corresponding .py files.

The Yelp and Amazon datasets will be automatically downloaded from the Internet.

Train BWGNN (homo) on Amazon (40%):

python main.py --dataset amazon --train_ratio 0.4 --hid_dim 64 \
--order 2 --homo 1 --epoch 100 --run 1

amazon can be replaced by other datasets: yelp/tfinance/tsocial

Train BWGNN (hetero) on Yelp (1%):

python main.py --dataset yelp --train_ratio 0.01 --hid_dim 64 \
--order 2 --homo 0 --epoch 100 --run 1

BWGNN (hetero) only supports Yelp and Amazon.

Train BWGNN (homo) on T-Social (40%):

python main.py --dataset tsocial --train_ratio 0.4 --hid_dim 10 \
--order 5 --homo 1 --epoch 100 --run 1

If you use this package and find it useful, please cite our ICML paper using the following BibTeX. Thanks! :)

@InProceedings{tang2022rethinking,
  title = 	 {Rethinking Graph Neural Networks for Anomaly Detection},
  author =       {Tang, Jianheng and Li, Jiajin and Gao, Ziqi and Li, Jia},
  booktitle = 	 {International Conference on Machine Learning},
  year = 	 {2022},
}

You can find a more detailed BibTex or other citation formats here.