MMSE: A Multi-Model Stacking Ensemble Learning Algorithm for Purchase Prediction

Analyzing users' consumption habits and preferences with big data technology, merchants can collect valuable data about users' consumption habits which is significant for them to make decisions, with the purpose of improving profitability and optimizing user experience. In this paper, we propose a Multi-Model Stacking Ensemble (MMSE) algorithm to deal with the problem of personalized commodities recommendation. The main components of our algorithm are the data analysis and the model construction. As for the data analysis, we apply data processing technologies to construct the data features. Specifically, we put forward a novel feature model which involves 6 clusters of features, and design a sampling algorithm to balance the ratio of positive and negative samples by the k-means clustering and undersampling. As for the model construction, we establish a two-layer model. In the first layer, we train four different ensemble algorithms as the basic classifiers, and in the second layer, we take XGBoost algorithm as the combiner classifier. Finally, we evaluate the proposed model on the users-commodities behavior data of Alibaba's M-Commerce platform. It performs better than single ensemble learning algorithms, achieving the performance of F1-score 10.97%.