Construction And Application Of Predictive Model Of User Purchase Behavior In The "New Retail" Scenario

With the maturity of information technology,mobile e-commerce,mobile payment,virtual reality and Io T smart terminals are widely used in social life,laying the foundation for the transformation of the retail industry.At the same time,due to the declining growth rate of Internet users and the gradual shrinking of traffic dividends,as well as factors such as labor costs and rent increases,the development of online e-commerce and offline traditional physical retail has encountered "bottleneck"."New retail",as a new business model of the retail industry,was born under this background.This paper uses a "new retail" platform as the background to study the purchase behavior of users in the category of shops,build a predictive model of user purchase behavior,explore more effective category strategy management methods,and help the transformation and upgrading of the retail industry.Based on the user behavior data of the platform,this paper first uses Neo4 j to construct a knowledge graph,constructs similar user sets for different users,performs data preprocessing for outliers and missing values,and conducts preliminary exploratory analysis on the data set.Secondly,it constructs its attribute characteristics,behavior characteristics,interaction characteristics and relationship characteristics centering on users,categories and shops,and selects the characteristics of the 186 characteristics.Third,construct XGBoost,LightGBM and CatBoost models respectively,and perform optimization and training.The accuracy and AUC of the three models are all greater than 0.9,and good results have been achieved.However,the improvement in F1-score is limited,all less than 0.66.In view of the shortcomings of the single model after optimization,this paper further studies the relationship between sample coverage and positive sample prediction accuracy under different models.The research finds that the accuracy rates of the top 20% of the XGBoost model and the top 25% of the positive sample predictions of the LightGBM model and CatBoost model are all greater than 0.9.Then,based on the cascade principle,analyze and compare the prediction effects of different primary and secondary models,set the CatBoost model as the primary model,XGBoost as the secondary model I and LightGBM as the secondary model II,and integrate the three models into a cascaded CXL-Boost model,While keeping Accuracy and AUC achieving high values,F1-score has been improved,and its value finally reached 0.7214.Finally,this paper is based on the importance and direction of the SHAP value research feature's influence on the model prediction results under the SHAP framework.The research results show that in terms of model performance,based on the multi-dimensional characteristics of users,categories,shops,etc.,the cascaded combination model can achieve better prediction effects than a single model in the prediction of“user-category-shop”.In feature contribution above,the relationship characteristics have a high degree of contribution to the prediction results of the model as a whole,and starting from similar user groups will help improve the management efficiency of the "new retail" platform.