BCAD: An Interpretable Anomaly Transaction Detection System Based on Behavior Consistency

Jun Hu, Xu Min, Xiaolu Zhang, Chilin Fu, Weichang Wu, Jun Zhou

Abstract: In the era of digital payment, abnormal behaviors such as fraud pose a huge threat to E-commerce platforms. Traditional anti-fraud approaches usually apply supervised learning which requires sophisticated knowledge extraction and does not adapt to evolving anomalous behaviors. In recent years, unsupervised learning methods has been widely applied to anomaly detection. However, they still suffer from a serious shortcoming, that they judge anomalies based on the global distribution while ignoring the user’s own historical information. In this paper, we propose a novel problem of unsupervised anomaly transaction detection focusing on the individual level. To tackle this problem, we first derive behavior consistency hypothesis based on data exploration. Then based on this assumption, we propose a new framework named Behavior Consistency based Anomaly Detection (BCAD). Specifically, BCAD learns representations for the target behavior and the history behavior preferences respectively by contrastive learning, and then measure the similarity between them to identify anomaly transactions. Besides, to disentangle the behavior representation into several attributes, we design an attribute gate module which can extract high-level user preferences from historical behaviors. Overall, BCAD can not only detect whether a target behavior is abnormal, even if the fraudulent pattern never appeared before, but also give an interpretation from the perspective of preference attributes. Extensive experiments on the real-world business dataset demonstrate that BCAD can detect abnormal behaviors effectively and provide insightful results for human beings.