Interpretable Graph-Level Anomaly Detection via Contrast with Normal Prototypes
Qiuran Zhao, Kai Ming Ting, Xinpeng Li
TL;DR
ProtoGLAD tackles graph level anomaly detection by grounding decisions in concrete normal graphs discovered from data via a point set kernel. It builds IK enhanced Weisfeiler Lehman embeddings, then grows normal clusters from prototypes and scores anomalies by their affinity to the nearest cluster, enabling explicit contrastive explanations. The approach yields competitive AUC performance across eight real datasets while providing interpretable node level cues that reveal which substructures drive anomalous decisions. This work enhances reliability and verifiability of GLAD in practical settings by tying anomalies to real normal exemplars rather than latent abstractions.
Abstract
The task of graph-level anomaly detection (GLAD) is to identify anomalous graphs that deviate significantly from the majority of graphs in a dataset. While deep GLAD methods have shown promising performance, their black-box nature limits their reliability and deployment in real-world applications. Although some recent methods have made attempts to provide explanations for anomaly detection results, they either provide explanations without referencing normal graphs, or rely on abstract latent vectors as prototypes rather than concrete graphs from the dataset. To address these limitations, we propose Prototype-based Graph-Level Anomaly Detection (ProtoGLAD), an interpretable unsupervised framework that provides explanation for each detected anomaly by explicitly contrasting with its nearest normal prototype graph. It employs a point-set kernel to iteratively discover multiple normal prototype graphs and their associated clusters from the dataset, then identifying graphs distant from all discovered normal clusters as anomalies. Extensive experiments on multiple real-world datasets demonstrate that ProtoGLAD achieves competitive anomaly detection performance compared to state-of-the-art GLAD methods while providing better human-interpretable prototype-based explanations.