Deep Learning for Anomaly Detection in Dynamic Graphs: A Verified Taxonomy, Survey, and Unified Benchmark

Anomaly detection in dynamic graphs (DGAD) sits at the intersection of graph learning and temporal modeling, yet it presents itself as a scattered collection of isolated methods: results are dispersed across dozens of venues, published numbers are mutually incomparable, and no reference delimits the field, so surveys routinely conflate dynamic-graph detectors with static- graph and multivariate time-series methods. This paper supplies the missing reference in both of its layers, organization and measurement. We delimit the field with three explicit, checkable inclusion criteria; we present a catalog of twenty-four methods, each traced to its primary publication and its publication status stated, organized by a two-axis taxonomy of structural encoder × temporal mechanism; and we list, with reasons, the prominent methods that fail each criterion. We then benchmark represen- tatives of all seven taxonomy families, thirteen deep detectors, a non-deep streaming baseline, and a suite of parameter-free heuristics, under one injection protocol, split, and metric set. The results concern the field’s evaluation practice as much as its methods. A degree heuristic that requires no training reaches AUC 0.811 on the canonical injected benchmark, above nine of thirteen deep detectors, and holds across a second corpus, a second injection model, and all 162 configurations of the protocol’s parameter space; on real fraud labels it falls below chance. There the ranking rearranges by time model rather than by injected rank: both snapshot-based leaders drop to chance while the continuous-time streaming detector leads. We further find that headline AUC misranks methods on injected data, and that published implementations frequently contain result- changing evaluation defects. We distill these findings into concrete recommendations and release the full benchmark