Predicting User Purchase Behavior via Multidimensional Feature Engineering and Explainable Machine Learning

Accurately predicting user purchase intentions is a critical challenge in e-commerce operations. Existing methods often rely on simple statistical features, lacking systematic theoretical guidance and model interpretability. To address these limitations, this study proposes a predictive framework that integrates the e-commerce conversion funnel model with interpretable machine learning. Utilizing a real-world dataset of approximately 37 million user behavior logs from Jingdong Group, we construct a five-dimensional feature engineering system comprising basic attributes, conversion rates, behavioral stability, overall activity, and interaction patterns. We evaluate four representative algorithms, including Decision Tree, Random Forest, XGBoost, and CatBoost. Experimental results demonstrate that the proposed multidimensional feature system significantly enhances the predictive performance of all tested models. Notably, CatBoost achieves the highest classification performance, reaching an accuracy of 0.9737 when utilizing all features. Furthermore, feature ablation experiments confirm the indispensable contribution of each extracted feature dimension. To overcome the black-box issue of machine learning models, we employ SHapley Additive exPlanations (SHAP) for in-depth feature attribution analysis. The SHAP results reveal that while basic traffic indicators establish a global baseline for purchase intent, behavioral stability and interaction features also serve as crucial leading indicators for final purchase decisions. This research provides a theory-driven paradigm for feature engineering and offers transparent insights for e-commerce platforms to optimize precision marketing strategies.