Residual-Augmented Latent Flow Matching for Graph Anomaly Detection

Unsupervised graph anomaly detection (GAD) aims to identify nodes that deviate from the majority of normal nodes without access to anomaly labels. This task remains challenging due to the scarcity of abnormal nodes and severe class imbalance. Graph neural network-based methods and recent diffusion-based generative models have made substantial progress in modeling complex graph data distributions. However, message passing can make anomalous node representations less distinguishable from their neighborhoods, while diffusion-based generative detectors often require costly multi-step denoising. We propose RALFlow-GAD, a residual-augmented latent flow matching framework for unsupervised graph anomaly detection. RALFlow-GAD augments graph autoencoder embeddings with degree-aware global and local residuals to preserve anomaly-relevant deviations, and learns a prototype-conditioned latent flow matching model to capture the dominant normal pattern. At inference time, RALFlow-GAD computes anomaly scores via one-step latent reconstruction and applies a lightweight score orientation correction. Experiments on five real-world datasets show that RALFlow-GAD achieves the best average performance among the reproduced methods. Compared with DiffGAD, the strongest reproduced baseline, RALFlow-GAD improves average AUROC by 10.3% and AP by 13.4%. Moreover, it obtains consistent speedups, achieving up to 41× over the diffusion-based generative baseline.