Prediction model of repurchase behavior of Tmall users and optimization of marketing strategy based on ensemble learning

Under the background that the focus of global e-commerce competition shifts to the value mining of existing users, this study constructs an ensemble learning model integrating interpretability technology to accurately predict repurchase behavior and optimize marketing strategy. Methodologically, this paper integrates multi-source e-commerce data, uses a weighted random forest algorithm optimized for sample imbalance, and incorporates SHAP value analysis for feature interpretation. The results show that the model achieves an AUC of 0.943 and an F1-Score of 0.816, outperforming baseline models. It demonstrates robustness (AUC decreases by only 3.75% under 20% noise) and practical value in A/B tests (conversion rate of 14.2%, ROI of 3.26). The SHAP interpretation aligns with expert judgment (88.8% consistency). The conclusion confirms the model's unity of prediction accuracy and interpretability, providing an end-to-end solution for e-commerce user lifecycle management.