Graph Anomaly Detection via Multi-Scale Reconstruction of Graph Encoder-Decoder Networks

Abstract

Existing unsupervised graph anomaly detection (GAD) methods can be categorized into reconstruction based methods and contrastive learning based methods. The principle of reconstruction methods is to capture anomalous nodes based on data reconstruction errors. However, existing reconstruction methods are limited to a single scale reconstruction of graph structures, typically predicting edges solely based on direct connections, without considering the rich contextual information within node neighborhoods. Moreover, reconstructing original low-density inputs usually introduce noise, where the reconstruction model may overfit the anomalous data. To this end, we propose a new reconstruction based graph anomaly detection method (GADMR for short). GADMR introduces a multi-scale reconstruction, which contains both structure and feature levels. From structure level, we use a node-subgraph reconstruction method to capture the relationship between a node and their neighboring subgraphs at multiple scales. From feature level, we employ a multi-layer representation reconstruction method to facilitate the evaluation of discrepancies in high-level feature representations. Extensive experiments on six benchmark datasets show that GADMR outperforms popular anomaly detection methods.