User Repeat Purchase Behavior Prediction Based on Accuracy-Weighted Stacking Algorithm

In this paper, an accuracy-weighted stacking fusion model based on Borderline-SMOTE is developed to predict the repeat purchase behavior of users on the Tmall platform.GBDT feature selection algorithm was utilized to obtain the importance ranking of features, filter redundant features, and construct a more efficient feature system.By calculating the error of the base learner's predicted results, weight coefficients were constructed to weigh the output.Subsequently, the new training set was used as input for the meta-learner.Through accuracy weighting, learners with better performance received higher weights, enabling weighted heterogeneous integration of the model and improving overall performance.The experiments verify that the accuracy weighting stacking algorithm proposed in this paper, based on Borderline-SMOTE, achieves high fitting accuracy.The AUC, F1, recall, and precision have all increased, with the AUC value reaching 0.93.The prediction effectiveness surpassed that of the XGBoost, LightGBM, CatBoost, Random Forest, and other models.This research will provide superior prediction models for e-commerce platform merchants, assisting them in capturing potential new users and retaining loyal customers, ultimately generating greater business value.