Wasserstein Heterogeneous Graph Neural Networks for Uncertainty-Aware Anomaly Detection

Abstract

Graph anomaly detection, a critical topic in graph mining, has garnered significant research interest and found applications across diverse domains such as attack event detection, spam review identification, and financial fraud prevention. Graph Neural Networks (GNNs) have emerged as the dominant approach in this field. Conventional GNN-based methods typically aggregate neighbor information to learn node embeddings and reconstruct structural relationships or attributes, assuming normal nodes exhibit lower reconstruction errors than anomalous ones. However, these methods often fail to account for uncertainty, higher-order structures, and graph heterogeneity, leading to suboptimal performance. To address these limitations, we propose a novel heterogeneous graph neural network that learns distribution-based node representations in Wasserstein space. Our approach leverages Gaussian distributions to capture uncertainty and employs Wasserstein distance to preserve transitivity, while incorporating reconstruction losses at structural, attribute, and type levels. Experimental results demonstrate that our proposed W-HGAD model achieves significant improvements over state-of-the-art methods, with AUC increases of 9.46% and 5.69% on two benchmark datasets. Ablation studies further validate the effectiveness of our model’s key components.