| In recent years,driven by emerging technologies such as big data and machine learning,effective analysis and prediction of consumer decisions have become a hot research topic.The growing richness of e-commerce environment has seriously impacted the offline real economy,but offline shopping,as a traditional consumption method,still has an important position in the digital era.This paper takes mobile phone as the research object to explore people’s behavioral preferences for offline shopping,aiming to help corporate brand stores better understand and serve the offline shopping population in the post-epidemic era and increase daily activity.In this paper,we use real user data from a cell phone brand operator after desensitization to conduct an in-depth study of their behaviors and basic characteristics.First,data visualization and analysis are conducted from three aspects:basic user characteristics,online tendencies and offline behaviors.Then,using feature engineering and feature screening methods,the data were pre-processed in three aspects:offline transaction preference,cell phone category preference,and operating brand preference,and finally 268 features were selected for subsequent empirical analysis.On this basis,the user purchase behavior prediction models of Random Forest.GBDT and LightGBM algorithms were constructed.Empirical findings:(1)In terms of overall accuracy,the three models performed comparably,all successfully mined the information in the data,the models had larger AUC values,and showed good applicability for predicting users’ online and offline purchase behavior;(2)Comparing the three selected empirical models,the LightGBM model required more parameters to be adjusted compared to the Random Forest and GBDT models,but it took less time overall.The LightGBM model focuses more on the basic information characteristics of users.Given that the classification and generalization ability of the single integrated learning model can be further improved,LightGBM is chosen as the improved model in this paper.The undersampling algorithms BalancedBaggingClassifier and EasyEnsembleClassifier are further combined with LightGBM to form the BBC-LightGBM model and EEC-LightGBM model,respectively.The overlapping validation on real data reveals that the improved models are significantly improved in terms of AUC values and recall rates.Among them,the improved EEC-LightGBM model can be better applied to offline purchase user prediction.For the problem of decreasing F1_score of the model,the observation is combined with the confusion matrix.The results show that the LightGBM after model fusion has significantly improved in predicting the accuracy of actual offline purchase users.i.e.,the recall rate has been effectively improved.Therefore,the improved LightGBM algorithm is worth trying in practical applications.Finally,for the prediction results,for both brands and operators,this paper puts forward some reasonable suggestions. |
