SAWGAD

Official implementation of Learning Feature Encoder with Synthetic Anomalies for Weakly Supervised Graph Anomaly Detection (IEEE TKDE 2026).

Requirements

• uv package manager
• CUDA 11.8 compatible GPU

Setup

1. Install uv.

2. Install dependencies:

uv sync

Usage

uv run sawgad

The default hyperparameters correspond to the configuration used in the paper for the Amazon dataset. The Amazon graph is downloaded automatically on first run via dgl.data.FraudAmazonDataset.

Expected Output

AUROC: 0.9702 +/- 0.0040 | AUPRC: 0.9418 +/- 0.0096 over 10 runs

Citation

If you find this code useful, please cite our paper:

@ARTICLE{Zhou26SAWGAD,
  author={Zhou, Yingjie and Xie, Yuqin and Liu, Fanxing and Song, Dongjin and Zhu, Ce and Liu, Lingqiao},
  journal={IEEE Transactions on Knowledge \& Data Engineering},
  title={{Learning Feature Encoder With Synthetic Anomalies for Weakly Supervised Graph Anomaly Detection}},
  year={2026},
  volume={38},
  number={04},
  ISSN={1558-2191},
  pages={2326-2339},
  doi={10.1109/TKDE.2026.3656821},
  url={https://doi.ieeecomputersociety.org/10.1109/TKDE.2026.3656821},
  publisher={IEEE Computer Society},
  month=apr
}