EVIDENT: An Ante-Hoc Evidence Bottleneck for Intrinsically Explainable Anomaly Detection in Dynamic Graphs

Abstract

Anomaly detection in dynamic graphs underpins fraud analysis, cybersecurity and platform integrity, yet deep detectors remain opaque: a flagged interaction arrives with a score and no account of which counterparty or which moment produced it. The prevailing post-hoc remedy fits an auxiliary model to a frozen detector, so faithfulness is estimated rather than guaranteed, and to our knowledge no dynamic-graph detector emits an explanation as a native product of its own forward pass. We propose EVIDENT (EVIDence-bottlenecked intrinsic dEtection for evolving Networks over Time), an ante-hoc detector whose anomaly score is computed exclusively from a sparse, dually-factorised evidence set selected within that pass, so the explanation costs no additional inference. Gated tokens leave the attention softmax entirely rather than being attenuated, so sufficiency holds by construction rather than by measurement. On Bitcoin-OTC with real distrust labels, EVIDENT attains 0.7947±0.0069 AUC under supervision matched to unsupervised baselines—exceeding the transformer detector TADDY by 26.0 points—from 10% of its evidence pool, retaining over 97% of an unconstrained model of identical architecture. Under a density-matched mask protocol its rationale exceeds an equalsize random subset by +0.2145 AUC (6.6σ) and selects priornegative-rating tokens at 1.29–2.49× the random rate; ablating the complement objective leaves detection unchanged while collapsing that advantage 4.5-fold. On a unified injected benchmark against eleven learned detectors and a parameter-free degree heuristic, EVIDENT ranks first at 0.9766 AUC while denied the cardinality channel they retain—though there its rationale ceases to separate, isolating token heterogeneity rather than the selection mechanism as the binding constraint on intrinsic explainability. We further document two evaluation defects—neighbourhood cardinality leakage under edge injection, and incomplete token removal in gated attention—with diagnostics that detect them.