Raising the Bar in Graph-level Anomaly Detection

TL;DR

This paper introduces a novel deep learning method for graph-level anomaly detection that leverages self-supervised and transformation learning techniques, significantly outperforming existing approaches on multiple real-world datasets.

Contribution

It proposes a new approach that addresses key issues in deep one-class methods for graph anomaly detection, improving accuracy and robustness.

Findings

Achieves 11.8% average AUC improvement over previous methods

Effectively addresses hypersphere collapse and performance flip issues

Demonstrates superior performance on nine real-world datasets

Abstract

Graph-level anomaly detection has become a critical topic in diverse areas, such as financial fraud detection and detecting anomalous activities in social networks. While most research has focused on anomaly detection for visual data such as images, where high detection accuracies have been obtained, existing deep learning approaches for graphs currently show considerably worse performance. This paper raises the bar on graph-level anomaly detection, i.e., the task of detecting abnormal graphs in a set of graphs. By drawing on ideas from self-supervised learning and transformation learning, we present a new deep learning approach that significantly improves existing deep one-class approaches by fixing some of their known problems, including hypersphere collapse and performance flip. Experiments on nine real-world data sets involving nine techniques reveal that our method achieves an average performance improvement of 11.8% AUC compared to the best existing approach.