Predicting User Purchase Behavior on JD.com: A Sequential Interaction and Engagement Depth Model

This study utilized transaction-level data from JD.com, as compiled for the 2020 MSOM Data Driven Research Challenge. By analyzing user click and purchase records along with associated product information, diversified classification models were introduced to predict user purchasing behavior. Following rigorous data preprocessing, the study implemented feature engineering based on user behavioral patterns, product attributes, and the interaction data between users and products. Specifically, through hierarchical modeling of user behavior, key features such as total engagement volume, maximum engagement depth, and click-through conversion rate were extracted and constructed to capture users' information acquisition efficiency and purchase intentions effectively. These features accurately reflect the sequential model of user behavior from browsing and clicking through to interactive requests, collectively contributing significantly to the optimization of both the accuracy and complexity of the classification model. The findings indicate that the Neural Network Classifier model outperforms others in predicting user ordering behavior, achieving an accuracy of 85.38%. This research offers a novel perspective for enhancing the performance and user experience of recommendation systems.