Label-based Graph Augmentation with Metapath for Graph Anomaly Detection

Abstract

Message-passing mechanism in GNNs could make the important characteristics for anomaly detection to become indistinguishable. How can we effectively obtain these characteristics even after message-passing for more effective graph anomaly detection? In this paper, we propose INFOREP, a framework for anomaly detection in attributed graphs. INFOREP consists of a novel graph sampling approach (INFOREP-S) and dual encoders. INFOREP-S enhances normality- and abnormality-specific information. The dual encoders effectively capture this enhanced information as well as complex interactions between normal and abnormal nodes. These processes allow us to grasp highly informative representations both locally and globally. INFOREP is (a) Accurate and fast: winning the highest average AUC and up to 20.1 faster than state-of-the-art approaches, (b) Scalable: linear in sampling subgraphs in input graph, processing millions of edges within 2 seconds, (c) Interpretable: visually explaining node representations, and (d) Robust: providing performance guarantees of real-world benchmarks with varying hyperparameters. Extensive experiments on six real-world benchmarks show that INFOREP achieves the highest average AUC value and the best performance on most datasets. Moreover, its node representations are visually well-separated as it preserves the important features.