New Benchmark Exposes GAD Model Failures on Million-Node Graphs

A newly introduced multi-dimensional benchmark for Graph Anomaly Detection (GAD) indicates that the majority of GNN-based techniques struggle to accommodate graphs with millions of nodes due to excessive memory demands. This benchmark, detailed in arXiv paper 2605.07133, rigorously assesses nine notable GAD models against three relevant challenges: graphs with millions of nodes, severe anomaly scarcity, and absent node attributes. Controlled variants are derived from five distinct graphs, which include two industrial-scale datasets exceeding 3.7 million nodes. The research reveals a significant disparity between academic assessments and practical applications, as current benchmarks focus on small, curated graphs with fairly balanced anomaly distributions. Findings emphasize that many methods falter under conditions of extreme anomaly scarcity and missing attributes, highlighting serious limitations for uses in financial fraud detection and governance on social platforms.