Domain adaptation for anomaly detection on heterogeneous graphs in e-commerce

Anomaly detection models have been the indispensable infrastructure of e-commerce platforms. However, existing anomaly detection models on e-commerce platforms face the challenges of “cold-start” and heterogeneous graphs which contain multiple types of nodes and edges. The scarcity of labeled anomalous training samples on heterogeneous graphs hinders the training of reliable models for anomaly detection. Although recent work has made great efforts on using domain adaptation to share knowledge between similar domains, none of them considers the problem of domain adaptation between heterogeneous graphs. To this end, we propose a Domain Adaptation method for heterogeneous GRaph Anomaly Detection in E-commerce (DAGrade). Specifically, DAGrade is designed as a domain adaptation approach to transfer our knowledge of anomalous patterns from label-rich source domains to target domains without labels. We apply a heterogeneous graph attention neural network to model complex heterogeneous graphs collected from e-commerce platforms and use an adversarial training strategy to ensure that the generated node vectors of each domain lay in the common vector space. Experiments on real-life datasets show that our method is capable of transferring knowledge across different domains and achieves satisfactory results for online deployment.