New Explainability Framework for Dynamic Graph Anomaly Detection

A new research paper introduces X-AddGraph, a novel framework designed to bring explainability to AddGraph, a prominent deep learning model for detecting anomalies in dynamic graphs. While AddGraph achieves high accuracy, its "black box" nature has hindered its adoption in regulated environments where auditability is crucial. X-AddGraph addresses this by providing clear reasons when an edge is flagged as anomalous. The core of X-AddGraph is its Dual Spatial-Temporal Attribution (DSTA) mechanism, which is specifically aligned with AddGraph's architectural components. This mechanism includes a gradient-based attribution for spatial relevance, a direct use of existing contextual attention weights for short-term temporal explanations, and a gradient rollback through recurrent hidden states for long-term temporal insights. Crucially, X-AddGraph is strictly post-hoc, meaning it does not alter the original detector's performance, maintaining identical detection scores. Empirical validation on the UCI Message benchmark shows that X-AddGraph effectively identifies historical snapshots with significant counterfactual signals, a capability lacking in spatially-blind explainers, thus enhancing trust and auditability in automated anomaly detection systems.