Anomaly Detection in government Data Graph Based on Channel and Causal Gating

In order to solve the problem of anomaly detection in the whole life cycle of government data, such as multi-source heterogeneity and high-dimensional noise, we propose the CAGAD two-stage graph anomaly detection framework. In the first stage, a lightweight channel attention module is used to dynamically suppress redundant features such as invalid user attributes and enhance local anomaly mining. In the second stage, a causal gated hybrid attention mechanism is designed to model the node-neighbor causal dependency of cross-departmental data sharing, and dynamically fuse the difference of abnormal scores ,feature similarity and causal signals. Experiments show that CAGAD achieves the best results on multiple benchmark data sets, and provides an efficient and interpretable solution for fraud detection and data tampering warning in government services.