Prediction Of User Repurchase Behavior Based On Machine Learning Methods

With the rapid development and popularization of the Internet,the scale of online shoppers is expanding rapidly and e-commerce is developing rapidly,which also leads to intensified competition among major e-commerce companies.Electric chamber of commerce in recent years to promote activities,such as a double tenth,618,issue coupons to attract new customers,but the effect of such activities is limited and no returns in the long run,because many consumers are just one-time customers.In order to obtain higher returns,businesses need to work out what kind of consumer repeat buyers is worth to pay.Therefore,the study of consumer repurchase behavior has a strong application value.This paper explores the re-purchase behavior of the website based on the user behavior record data of "Double 11" in 2017.The data included the information of some merchants' new customers on the activity day of November 11 and in the 6months before.Firstly,from the perspective of business,this paper designed feature engineering to extract features by analyzing the factors affecting repeat purchase,and used three models for modeling and the corresponding model fusion.The algorithm performed well in the experiment.The work of this paper mainly includes the following aspects:(1)Data processing.Through data introduction,data exploration and data preprocessing,we can better understand the field meaning of data set and its relationship with whether the label is repurchased or not.(2)Design feature engineering.By analyzing the factors affecting repeat purchase behavior,the paper extracts the features from three perspectives: user characteristics,merchant characteristics,and the interaction between users and merchants.Feature selection is carried out after dimension increase.(3)Study the application of single model and fusion model in feature engineering.XGB,LGB and Catboost were used for modeling respectively.The last several models were combined together,and the combined model could complement the advantages of each model to obtain better prediction results.