TUAF: Triple-Unit-Based Graph-Level Anomaly Detection with Adaptive Fusion Readout

Zhenyang Yu, Xinye Wang, Bingzhe Zhang, Zhaohang Luo, Lei Duan

Abstract: Graph-level anomaly detection (GAD) has emerged as a significant research direction due to its practical application in diverse domains, such as toxic drug identification and compound activity assay. Existing GAD methods generally regard the node as the basic unit to learn graph representation, thus ignoring the vital information of the triple structure “node-edge-node". Intuitively, the occurrences of anomalous events in form of triples are the primary cause of an abnormal graph. Meanwhile, previous works adopt trivial readout strategies to obtain the graph-level representation without considering the different contributions of nodes. In this paper, we propose a novel GAD method named TUAF, based on triple-unit graphs with adaptive fusion readout. Specifically, we first transform the original graph into the triple-unit graph, and then learn triple representations for capturing abundant information about an edge and its corresponding nodes simultaneously. Furthermore, we design an adaptive fusion readout to obtain a high-quality graph-level representation by adaptively learning the optimal gravity coefficient for each triple. Through extensive experiments, we demonstrate the effectiveness of TUAF on 18 real-world datasets.