X-AddGraph explains graph anomaly AI with zero AUC loss

Dynamic graph anomaly detectors flag suspicious edges in evolving networks—like unusual communications in enterprise systems—but typically output only a score, leaving analysts without a reason for the alert. That opacity is a major issue in regulated, cooperative information systems where automated decisions must be auditable. Now, researchers from the team of Iyad Assaad Nekka have released X-AddGraph, the first explainability framework for AddGraph, a foundational GCN+GRU model for edge-level anomaly detection. X-AddGraph presents a Dual Spatial-Temporal Attribution (DSTA) mechanism with three modules, each aligned to AddGraph's architecture: gradient-based relevance over the current adjacency structure for spatial explainability, direct reading of contextual attention weights computed during inference for short-term temporal insight (at zero additional cost), and gradient rollback through recurrent hidden states for long-term temporal analysis.

Because X-AddGraph is strictly post-hoc, the frozen AddGraph detector preserves its performance exactly—the paper verifies Delta AUC = 0 empirically to ten decimal places. On the UCI Message benchmark, the baseline AddGraph reaches an average per-snapshot AUC of 0.8705, exceeding originally published results, and X-AddGraph reproduces every score identically while adding explanations. Evaluated across confident true positives, low-confidence positives, false positives, and random edge samples, the long-term attribution component identifies historical snapshots with significantly more counterfactual signal than random selection (0.127 vs 0.074)—capability no spatially-blind explainer can offer. The authors plan to release their implementation for full reproducibility, targeting CoopIS 2026.