Rethinking Graph Neural Networks for Anomaly Detection (github.com)

The authors of the ICML 2022 paper "Rethinking Graph Neural Networks for Anomaly Detection" released the official implementation of their BWGNN model and integrated it into GADBench, a comprehensive (semi-)supervised benchmark for graph anomaly detection. This makes the paper’s claims reproducible and directly comparable to other methods on standardized tasks, which matters for researchers evaluating detection performance across homogeneous and heterogeneous graphs. The repo also provides the T-Finance and T-Social datasets used in the paper (plus automatic download for Yelp and Amazon) and a plot.zip to reproduce key figures. Practically, the codebase targets PyTorch 1.9.0 and DGL 0.8.1 (with sympy, argparse, sklearn), and exposes training scripts with typical hyperparameters: --hid_dim (hidden size), --order (aggregation depth), --homo (homogeneous vs. heterogeneous), --train_ratio, --epoch and --run. Example commands show BWGNN(homo) on Amazon/T-Social and BWGNN(hetero) on Yelp/Amazon (note: hetero mode supports only Yelp and Amazon). The package therefore facilitates hands-on evaluation, ablation, and benchmarking of BWGNN in both research and applied settings; users are asked to cite the ICML paper for academic use.