Deep Graph Anomaly Detection: A Survey and New Perspectives (github.com)

A comprehensive survey paper, "Deep Graph Anomaly Detection: A Survey and New Perspectives" (TKDE, 2025), plus a professionally curated “awesome” repository of papers, code and datasets, has been released—accompanied by a half‑day tutorial on August 18, 2025 (slides available). This is presented as the first systematic consolidation of recent DGAD work, organized into a clear taxonomy (GNN backbone design, proxy task design, anomaly measures, datasets, quantitative comparisons and related surveys). The resource list spans key advances across NeurIPS/ICLR/KDD/WebConf/ICML and ArXiv, and the authors will keep it updated and welcome community contributions. Technically, the survey synthesizes recurring solutions and pain points: contrastive and self‑supervised proxy tasks, reconstruction vs. generative augmentation, handling heterophily and class imbalance, dynamic/federated settings, few‑shot and pre‑training approaches, and even training‑free SVD and LLM‑assisted methods. The quantitative tables show wide AUROC variability across benchmarks (Cora, Citeseer, Amazon, Elliptic, financial/fraud datasets), highlighting inconsistent evaluation protocols and persistent challenges—OOD generalization, label scarcity, privacy-preserving/federated detection, and real‑world fraud heterogeneity. For practitioners and researchers, the survey is a practical roadmap for method selection, reproducible benchmarking, and open problems that should guide next‑generation DGAD work.