This is the code associated with the submission "Boosting Graph Anomaly Detection with Adaptive Message Passing".

conda install pytorch==1.12.0 torchvision==0.13.0 torchaudio==0.12.0 cudatoolkit=11.6 -c pytorch -c conda-forge
conda install -c dglteam/label/cu116 dgl
conda install scikit-learn
pip install pygod

We provide five benchmark datasets containing injected anomalies: Cora, Citeseer, Pubmed, ACM and BlogCatalog, as well as two real-world datasets containing organic anomalies: books and reddit, which can be found in data/data.rar. Anomalies are injected through the unified interface provided by the pygod library. Two OGB datasets ogbn-arxiv and ogbn-products are not included due to memory limits.

We recommend using preprocessed data for fair comparasion, unzip data/data.rar and make directory structure as follows:

└─data
    │      Cora.bin
    │      Citeseer.bin
    │      ...
└─run.py

For two OGB datasets or customized dataset, contextual and structural anomalies can be generated via 'pygod.generator'. See https://docs.pygod.org/en/latest/pygod.generator.html for details.

Run python run.py --data Cora --local-lr 1e-3 --local-epochs 100 --global-lr 5e-4 --global-epochs 50 to perform anomaly detection.