Post-Hoc Explainability for Contrastive Graph Anomaly Detection via Gradient Attention Maps
Year: 2026
Type: preprint
Abstract: Graph anomaly detection on attributed networks has seen significant advances through deep learning, with contrastive self-supervised methods such as CoLA achieving stateof-the-art performance. However, these methods operate as black boxes: they flag nodes as anomalous without providing any explanation of which structural connections or feature dimensions drove that decision. In operational settings such as financial fraud detection, cybersecurity incident response, and clinical surveillance, a detection system that cannot explain its decisions is difficult to trust, audit, or act upon. In this paper, we introduce X-CoLA, the first post-hoc explainability framework for contrastive graph anomaly detection. X-CoLA integrates Gradient Attention Maps (GRAM) into the CoLA framework, applying a single backward pass through CoLA’s trained GCN backbone to produce node-level gradient attention maps that identify, for each flagged anomaly, which neighbors in the local subgraph and which feature dimensions of the target node drove the contrastive disagreement score. X-CoLA simultaneously distinguishes between structural anomalies, explained by high-gradient edges, and contextual anomalies, explained by high-gradient feature dimensions, without any modification to CoLA’s training procedure. We define three explainability metrics adapted to the contrastive setting: Gradient Fidelity, Structural Sufficiency, and Contextual Comprehensiveness. Experiments on the Cora benchmark dataset demonstrate that XCoLA achieves a Structural Sufficiency of 0.906 and a Gradient Fidelity of 0.602 while preserving CoLA’s detection AUC exactly. To the best of our knowledge, X-CoLA is the first work to apply post-hoc gradient-based explainability to a contrastive selfsupervised graph anomaly detection method.
Source: Research Square
Institutions: École Nationale Supérieure d'Informatique, Lyon 1 Université
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7204236520