Research On Prediction Of User Buying Behavior Based On Data Balance And Model Fusion

The rapid development of e-commerce has brought great convenience to people's life,but the variety of its products makes users need to spend more time and energy to find the right products,so understanding the user's purchase intention is the key to improve the user's shopping experience on e-commerce platform The research focus of this paper is to build a two classification prediction model to predict the purchase of users.The main work of the paper is as follows:(1)Construct predictive characteristics of users' purchasing behavior.Building predictive features is an important part of building predictive models.Good features can be used with common algorithms to achieve excellent predictive results.First extract basic features from the original data,then use statistical knowledge to construct some complex derivative features,and finally analyze the correlation of features to remove irrelevant predicted features.(2)Improve the random under-sampling data balance method.In the user's purchase behavior data,the user's browsing and shopping cart behavior data are much more than the purchase behavior data.This paper proposes an improved random under-sampling processing imbalance problem based on the K-means algorithm.Improved random under-sampling utilizes K-means to cluster the majority of samples and then delete the samples from each cluster,effectively solving the problem of information loss in the random under-sampling method.(3)Multi-heterogeneous algorithm fusion prediction model.The advantages of the fusion model collection of multiple prediction algorithms can obtain better prediction results.This paper combines the timing of the long and short-term memory and the extreme gradient boosting for the generalization of sparse data and logistic regression for the robustness of small and medium noises,the three algorithms are fused using the Stacking ensemble learning method to obtain a prediction model to predict user purchasing behavior.We uses the experimental data provided by JD big data competition to verify the prediction effect of the model.The experimental results show that the improved random under sampling data balance method can effectively prevent the loss of data set information and achieve the effect of data balance.The prediction effect and generalization ability of the fusion prediction model is better than that of the single prediction model.