ChiGAD Uses Chi-Square Wavelets to Detect Anomalies in Heterogeneous Graphs
New spectral GNN beats state-of-the-art on multiple anomaly detection benchmarks.
Graph Anomaly Detection (GAD) in heterogeneous networks is challenging because nodes and edges have multiple types, requiring models to capture rich semantics and subtle abnormal signals. Existing GNN methods often fail to preserve high-frequency information and struggle with imbalanced anomaly samples. To overcome these, a team of researchers (led by Xiping Li) introduces ChiGAD, a spectral GNN framework built on a novel Chi-Square filter inspired by wavelet effectiveness. The framework has three core components: a Multi-Graph Chi-Square Filter that applies dedicated filters to each meta-path graph, an Interactive Meta-Graph Convolution that aligns features while retaining high-frequency content, and a Contribution-Informed Cross-Entropy Loss that prioritizes difficult anomalies to mitigate class imbalance.
Extensive experiments on both public and industrial datasets show ChiGAD significantly outperforms state-of-the-art models across multiple metrics. Additionally, the team's homogeneous variant, ChiGNN, achieves top results on seven standard GAD datasets, validating the power of Chi-Square filters for graph anomaly detection. The code is publicly available on arXiv. This work advances the ability to detect fraud, bot activity, and other subtle anomalies in complex relational networks where heterogeneity is the norm.
- ChiGAD introduces a Chi-Square wavelet filter to capture abnormal signals across diverse meta-paths in heterogeneous graphs.
- The framework uses Interactive Meta-Graph Convolution to preserve high-frequency information during feature alignment.
- Contribution-Informed Cross-Entropy Loss prioritizes difficult anomaly samples to address class imbalance, boosting detection accuracy.
Why It Matters
Fraud and anomaly detection in complex networks just got a powerful new tool that handles heterogeneity and imbalance better than ever.