Zero-Degradation Post-Hoc Explanations with Identity Pass-Through for Recurrent Graph Anomaly Detection
Year: 2026
Type: preprint
Abstract: Anomaly detection in dynamic graphs underpinshigh-stakes applications in financial fraud analysis, intrusiondetection, and online platform integrity, where accountabilityframeworks increasingly require that automated decisions beaccompanied by human-interpretable justifications. StrGNN, thestrongest performer in recent unified benchmarks of the field,delivers this detection quality while producing no explanationwhatsoever: when an edge is flagged, the analyst receives ascore and nothing more. Explanation quality metrics are notlow for StrGNN — they are undefined, because no attribu-tion vector exists to evaluate. This paper closes that gap. Wepresent X-StrGNN, a post-hoc explanation layer that wrapsa trained, frozen StrGNN and emits, for every flagged edge,two complementary attributions: a structural attribution overthe enclosing-subgraph evidence identifying which contextualinteractions drove the decision, and a temporal attribution overthe observation window identifying which historical snapshotcarried the signal. Because both attributions are realised asmultiplicative masks that are identically one in the unexplainedpass, the layer is an exact pass-through: detection is preservedto machine precision, which we verify rather than assert (∆AUC= 0.0000, ∆AP = 0.0000, ∆P@100 = 0.0000). Attributioncosts 0.66 ms per flagged edge, making explanation of an entirealarm list feasible rather than a hand-selected handful. Wefurther conduct the first controlled design study of attributionstrategies for this architecture, comparing gradient attribution,per-instance mask optimisation, and amortised parameterisationunder one protocol, one budget, and three seeds. X-StrGNNattains the highest attribution stability in the study (0.913) at268× lower cost than per-instance optimisation, and its temporalattribution (1.601 against a measured random floor of 0.973) isseparably better than its own ablated control, while per-instanceoptimisation — the most expensive strategy evaluated — fallsbelow that floor. Code, protocol, and per-seed measurements arereleased.
Source: HAL
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: W7204349929