Temporally Contextualized Supervision for Dynamic Graph Anomaly Detection

Dynamic graph anomaly detection aims to identify abnormal behaviors in evolving interaction streams. Existing methods often rely on point-level evidence scores, such as drift scores that capture abnormal temporal deviations in interaction patterns or node representations, and reconstruction scores that measure reconstruction inconsistency with historical graph states. Although such point-level evidence scores can characterize abnormality-related signals, using them as isolated training signals assumes that the evidence for an anomaly is temporally aligned with the observed label at the same interaction. In this work, we show that this assumption does not always hold in dynamic graphs: the evidence associated with an annotated anomaly may emerge before the labeled interaction, persist around it, or extend beyond it, leading to an evidence–label temporal mismatch. To address this issue, we propose TECOS, a temporally contextualized supervision framework for dynamic graph anomaly detection. Instead of applying supervision to isolated point-level evidence, TECOS uses short-range source-wise history together with source- and destination-side representations to learn a bounded correction to a reconstruction–drift base score, better aligning supervision with temporally misaligned anomaly evidence. We further provide empirical and theoretical analyses to characterize when point-level learning remains sufficient and when temporal context becomes beneficial. Experiments on seven temporal interaction datasets show that TECOS achieves the best overall performance across AUC and AP, demonstrating the effectiveness of temporally contextualized supervision.