This is the code for WSDM2022 paper "Deep Graph-level Anomaly Detection by Glocal Knowledge Distillation".

Global and Local Knowledge Distillation (GLocalKD) is introduced in our WSDM22 paper, which leverages a set of normal training data to perform end-to-end anomaly score learning for graph-level anomaly detection (GAD) problem. GAD describes the problem of detecting graphs that are abnormal in their structure and/or the features of their nodes, as compared to other graphs. GLocalKD addresses a semi-supervised GAD problem in that the data known are all labeled normal data. The experiment results show that GLocalKD can be implemented data-effectively and is robustness to anomaly contamination, indicating its applicability in both unsupervised (anomaly-contaminated unlabeled training data) and semi-supervised (exclusively normal training data) settings.

Some of datasets are put in ./dataset folder. Due to the large file size limitation, some datasets are not uploaded in this project. You may download them from the urls listed in the paper.

For datasets except HSE, p53, MMP, PPAR-gamma and hERG, run the following code. For datasets with node attributes, feature chooses default, otherwise deg-num.

python main.py --dataset [] --feature [default/deg-num]

For HSE, p53, MMP and PPAR-gamma, run the following code.

python main_Tox.py --dataset []

For hERG, run the following code.

python main_smiles.py

@inproceedings{ma2022deep,
  title={Deep Graph-level Anomaly Detection by Glocal Knowledge Distillation},
  author={Ma, Rongrong and Pang, Guansong and Chen, Ling and van den Hengel, Anton},
  booktitle={WSDM '22: The Fifteenth ACM International Conference on Web Search and Data Mining},
  year={2022}
}