Study On Prediction Of Repurchase Behavior Of Business Users Based On Classification Model

E-commerce websites serve hundreds of millions of users every day.While people are shopping on Taobao,Jingdong,Dangdang and other e-commerce platforms,they will produce huge amounts of data.It is of great significance to dig out potential value from them.Rational use of these data can bring users a better consumption experience.This thesis takes the real user,commodity and behavior data of a certain period after desensitization treatment of Jingdong Mall Platform as the research object,and combines data mining technology and machine learning classification algorithm.The main research work is as follows:(1)Characteristic Engineering Research on consumer behavior data of e-commerce users.Emphasis is laid on data preprocessing,feature analysis,feature construction and feature selection techniques.Users’ repurchase behavior characteristics are excavated,and statistical features,derivative features and behavior attenuation characteristics are constructed.The original data are transformed into high-dimensional and trainable effective sample data,and data features affecting the top 10 importance of users’ repurchase behavior are selected by combining correlation coefficient and random forest.(2)A comparative study of classification models.From the traditional user-commodity point of view,combined with the classification algorithm in machine learning,three prediction models are selected,which are logistic regression,Xgboost and support vector machine.The training parameters of the model are optimized by cross-validation and grid search.The comparison results show that the prediction effect of the model based on Xgboost algorithm is better than other models.(3)Research on model fusion method.Through the research on the idea and method of model fusion,this thesis uses the Soft-Voting method to achieve the fusion of prediction models.The experimental results show that the prediction accuracy of the fusion model is improved by about 3% compared with that before fusion.This thesis studies the data-driven prediction of the repurchase behavior of e-commerce users,analyses the construction characteristics from two perspectives ofuser-commodity and user-commodity categories,and proposes a scheme that integrates the user-commodity perspective(U-I)prediction model with the user-commodity category perspective(U-C)prediction model,which effectively solves the data sensitivity and ease of the single-angle model.The over-fitting problem can be applied to e-commerce business scenarios to help e-commerce platforms achieve accurate marketing and improve user retention.