Research map / Graph learning / Graph self-supervised learning

ARC: A Generalist Graph Anomaly Detector with In-Context Learning

NeurIPS 2024 · Graph learning › Graph self-supervised learning · approach: Graph anomaly detection · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes ARC, a generalist graph anomaly detector that is trained once on several graph datasets and detects anomalies on unseen datasets via in-context learning from few-shot normal nodes, without fine-tuning.

Problem

Graph anomaly detection methods need dataset-specific training, large in-domain data, and generalize poorly to new graph datasets with different feature spaces.

Approach

Aligns heterogeneous node features by projection plus smoothness-based feature sorting, encodes nodes with a multi-hop ego-neighbor residual GNN, and scores anomalies by how far a query embedding drifts when cross-attentively reconstructed from few-shot normal context nodes.

Key results

• ARC achieves best AUROC on 5 of 8 test datasets and average rank 1.5, e.g., 87.45 on Cora and 90.95 on CiteSeer.
• Gains over the best baseline reported as +21.1% on Cora, +18.8% on CiteSeer, and +36.6% on Amazon.
• Supervised and unsupervised pre-train-only baselines are near random (about 50% AUROC) on many datasets, while ARC needs no fine-tuning.

Measured effects reported in the paper

Surprising findings

• Pre-train-only GAD baselines often perform close to random guessing on unseen datasets.
• Dataset-specific fine-tuning sometimes yields minor or negative gains over pre-train-only baselines, suggesting negative transfer.

Limitations

• Can only use normal context samples at inference; cannot exploit labeled abnormal context samples.
• Performance is weaker on some datasets such as Weibo, where the GCN encoder beats the residual encoder in the ablation.
• Feature alignment relies on smoothness heuristics, which are empirically motivated rather than theoretically guaranteed.

Experimental setup

Datasets: PubMed Flickr Questions YelpChi Cora CiteSeer ACM BlogCatalog Facebook Weibo Reddit Amazon

Benchmarks: Cora CiteSeer ACM BlogCatalog Facebook Weibo Reddit Amazon

Metrics: AUROC AUPRC

Baselines: GCN GAT BGNN BWGNN GHRN DOMINANT CoLA HCM-A TAM

Builds on: TAM SGC BWGNN InCTRL PRODIGY

Closest papers by meaning

• ADA-GAD: Anomaly-Denoised Autoencoders for Graph Anomaly Detection (AAAI 2024 · similarity 0.92)
• Open-Set Graph Anomaly Detection via Normal Structure Regularisation (ICLR 2025 · similarity 0.91)
• Graph Anomaly Detection as Dynamical Transport: Training-Free Scoring via Empirical Bayes (NeurIPS 2026 · similarity 0.91)
• Generative Semi-supervised Graph Anomaly Detection (NeurIPS 2024 · similarity 0.91)
• UniFORM: Towards Unified Framework for Anomaly Detection on Graphs (AAAI 2025 · similarity 0.90)
• UniGAD: Unifying Multi-level Graph Anomaly Detection (NeurIPS 2024 · similarity 0.90)