While users consume and shop on e-commerce platforms, they will generate a huge amount of data information, and tapping the potential value of these data can optimize online marketing and bring users a better consumption experience. This study aims to predict users’ repurchase behavior and formulate personalized marketing strategies by analyzing their repurchase behavior on e-commerce platforms. First, the improved RFM model and K-means++ algorithm are utilized for user value classification. Then, a model for predicting user repurchase behavior was constructed based on Logistic regression, XGBoost, and SVM, respectively, and the prediction effects were compared. Then, the prediction models UI and U-C are built based on the XGBoost algorithm from the perspective of user and product category, respectively, and fused using the Soft-Voting method. The prediction effect of the fused models is verified at the end. The F1 values for all three models in the test set are approximately 0.2, and the XGBoost model has a significantly superior prediction effect than the other two models. The precision, recall, and F1 values of the fused model are about 0.31, 0.26, and 0.28, respectively. These values have been improved by about 4%-19% compared to the pre-fusion. The fusion model’s ROC curve is located at the upper left corner and has an AUC of 0.82, indicating high accuracy and stable results. This study provides feasible suggestions for the development of online marketing strategies to promote user repurchase behavior.