TGT: Churn Prediction in O2O Logistics with Two-tower Gated Transformer

Online-to-Offline (O2O) logistics has developed rapidly and provided great convenience to people’s lives. In O2O logistics, one of the essential problems is to predict the logistics customer churn behaviors. Traditional rule-based or survey-based churn prediction methods are laborious and time-intensive while existing machine learning-based churn prediction studies do not consider combining online and offline behaviors, rendering them impractical to be applied to O2O logistics. In this paper, we leverage historical online store operations, logistics provider selections, and the quality of logistics transportation data as the opportunity to investigate customer churn considering both online and offline behaviors. Specifically, we propose TGT, a novel Two-tower Gated Transformer framework for logistics customer churn prediction. TGT consists of two key modules: (i) a two-tower structure to extract complex dependencies among various variables over diverse periods; and (ii) a scale-adaptive convolution module (SAC) to capture the effect of various factors on customer churn in the O2O logistics system across different time scales. We evaluate the effectiveness of the TGT model based on more than half a year of real-world e-commerce data incorporating over 54,000 merchants and 877 million orders, collected from one of the largest logistics platforms in China. Experimental results indicate that our method achieves a 5.5% improvement in the F1 score prediction results. In addition, we deploy TGT at JD Logistics for the real-world churn prediction and reduce the churn rate from 11.75% to 5.94%, which verifies the effectiveness and impacts of the proposed method.