Chi-Square Wavelet Graph Neural Networks for Heterogeneous Graph Anomaly Detection
Organizations: The Chinese University of Hong Kong, Hong Kong SAR, China · Mohamed bin Zayed University of Artificial Intelligence, Abu Dhabi, United Arab Emirates · Tencent, WeChat Pay, Shenzhen, China
Abstract
Graph Anomaly Detection (GAD) in heterogeneous networks presents unique challenges due to node and edge heterogeneity. Existing Graph Neural Network (GNN) methods primarily focus on homogeneous GAD and thus fail to address three key issues: (C1) Capturing abnormal signal and rich semantics across diverse meta-paths; (C2) Retaining high-frequency content in HIN dimension alignment; and (C3) Learning effectively from difficult anomaly samples with class imbalance. To overcome these, we propose ChiGAD, a spectral GNN framework based on a novel Chi-Square filter, inspired by the wavelet effectiveness in diverse domains. Specifically, ChiGAD consists of: (1) Multi-Graph Chi-Square Filter, which captures anomalous information via applying dedicated Chi-Square filters to each meta-path graph; (2) Interactive Meta-Graph Convolution, which aligns features while preserving high-frequency information and incorporates heterogeneous messages by a unified Chi-Square Filter; and (3) Contribution-Informed Cross-Entropy Loss, which prioritizes difficult anomalies to address class imbalance. Extensive experiments on public and industrial datasets show that ChiGAD outperforms state-of-the-art models on multiple metrics. Additionally, its homogeneous variant, ChiGNN, excels on seven GAD datasets, validating the effectiveness of Chi-Square filters. Our code is available at https://github.com/HsipingLi/ChiGAD.
Figures & tables
| Metric | AUROC | AUPRC | F1-macro | Recall |
|---|---|---|---|---|
| LL (employed) | 0.9702 | 0.9518 | 0.9195 | 0.8947 |
| 0.6587 | 0.5035 | 0.5975 | 0.4687 | |
| 0.5746 | 0.3852 | 0.5393 | 0.3885 | |
| 0.5714 | 0.3854 | 0.4417 | 0.3784 | |
| 0.9597 | 0.9088 | 0.9157 | 0.8847 | |
| 0.9483 | 0.8928 | 0.8798 | 0.8296 |
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
| Dataset | #Node | #Edge | #Feat. | Anomaly % | Split % (Train/Val/Test) | Relation | Feat. Type |
| ACM (het) | 10,942 | 547,872 | [193,115,250,95] | 32.68/32.07/35.07 | 40/20/40 | Co-authorship | Misc. Info. |
| R-I (het) | 2,880,722 | 8,781,690 | [14,171,10,13] | 0.35 | 40/20/40 | Account Interaction | Misc. Info. |
| R-II (het) | 6,168,746 | 15,576,508 | [14,171,10,13] | 0.55 | 40/20/40 | Account Interaction | Misc. Info. |
| 10,984 | 168,016 | 64 | 3.3 | 40/20/40 | Under Same Post | Text Embedding | |
| Tolokers | 11,758 | 519,000 | 10 | 21.8 | 50/25/45 | Work Collaboration | Misc. Info. |
| Amazon | 11,944 | 4,398,392 | 25 | 9.5 | 70/10/20 | Review Correlation | Misc. Info. |
| Hyper-parameter | ACM | R-I | R-II | Tolokers | Amazon | T-Finance | Questions | DGraph-Fin | T-Social | |
| learning rate | 0.0001 | 0.006 | 0.006 | 0.0003 | 0.003 | 0.0003 | 0.003 | 0.0001 | 0.01 | 0.01 |
| dropout rate | 0 | 0 | 0 | 0 | 0.3 | 0.6 | 0.6 | 0.4 | 0.4 | 0 |
| hidden dimension | 512 | 25 | 48 | 64 | 64 | 64 | 80 | 400 | 32 | 20 |
| aligned dimension | 512 | 25 | 48 | - | - | - | - | - | - | - |
| epochs | 200 | 200 | 200 | 1700 | 700 | 1550 | 300 | 150 | 1200 | 4000 |
| MLP layers | 4 | 2 | 2 | 2 | 2 | 3 | 3 | 4 | 2 | 3 |
| 1 | 0.6970 | 0.0000 | 16 | 1.5940 | 1.7638 |
|---|---|---|---|---|---|
| 2 | 0.9603 | 0.6667 | 32 | 1.7126 | 1.8779 |
| 4 | 1.2180 | 1.1992 | 64 | 1.7973 | 1.9339 |
| 8 | 1.4313 | 1.5556 | 128 | 1.8571 | 1.9600 |
| Model | AUROC | AUPRC | F1-macro | Recall |
|---|---|---|---|---|
| ChiGAD | 0.9702 | 0.9518 | 0.9195 | 0.8947 |
| w/o Multi-Graph Chi-Square Filter | 0.9538 | 0.9236 | 0.8889 | 0.8499 |
| w/o Interactive Meta-Graph Chi-Square Convolution | 0.8969 | 0.8410 | 0.7988 | 0.7343 |
| w/o Contribution-Informed Cross-Entropy Loss | 0.9237 | 0.8449 | 0.8526 | 0.8070 |
| Variant of ChiGAD (Beta filter) | 0.9411 | 0.9093 | 0.8735 | 0.8371 |