RK-W-Stacking: A Hybrid Model Combining Entropy-Weighted RFM-K-Means Clustering and Weighted Stacking Ensemble for User Consumption Prediction

Abstract

Understanding user consumption characteristics helps predict consumer behavior and design appropriate marketing strategies, which promotes spending and supports market economic circulation. However, existing models have a low recognition rate for user consumption characteristics and limited accuracy in consumer behavior prediction. To this end, a user consumption behavior prediction model was constructed in the research. The main innovation of the research method lies in the weighted reconstruction of the K-means distance calculation logic by feature variance to enhance the robustness of the non-convex data set class. The index weights are dynamically calculated through the entropy weight method to optimize the user value assessment. A Stacking ensemble framework based on error rate weighting is designed to integrate the advantages of XGBoost and random forest-based classifiers. Experimental results showed 96.3% feature recognition accuracy, 0.25% loss rate, and 0.91 AUC, demonstrating strong classification ability. Additionally, long-term consumption prediction experiments show that the proposed model reaches a maximum memory usage of 1170 MB, a maximum response time of 120 ms, while maintains an annual prediction accuracy of over 94.2%. These results indicate that the proposed model achieves high accuracy and adaptability in extracting consumption characteristics and forecasting future consumer behavior. Furthermore, the system exhibits greater stability and faster response times than comparable prediction models, providing new insights for merchants to optimize business strategies and anticipate market trends.