Product Anomaly Detection on Heterogeneous Graphs with Sparse Labels

Dan Yin, Sihang Fang, Tianshuo Wang, Maozu Guo

Abstract: With the increasing popularity of online shopping platforms such as Amazon, Taobao and eBay, it is an urgent need to detect anomalous products, i.e., fake and illegal commodities. Heterogeneous graphs, which represent multiple types of objects and their relationships, have emerged as powerful and expressive representations for modeling real-world interactions in various domains like knowledge graphs, social networks and sales networks. In this study, we focus on detecting product anomalies on heterogeneous graphs. It is important to note that our model can be comprehensively applied to any anomaly detection application, not just limited to product anomalies. To tackle this problem, we encounter three noteworthy challenges: class imbalance, high heterogeneity, and label sparsity. However, existing approaches have limitations in effectively addressing these issues. To resolve the challenges, we propose a novel approach for product anomaly detection on heterogeneous graphs. Our approach consists of three key modules: 1) An imbalanced sample strategy that effectively handles class imbalance and high heterogeneity; 2) A label propagation module that tackles the issue of label sparsity; and 3) Jumping Knowledge Network (JK-net) for adaptively information aggregation. Extensive experiments demonstrate that our method outperforms state-of-the-art baselines on real world datasets by a significant margin. All source codes are available on https://github.com/FXiaoHang/anomalydetection.