Research On User Purchase Prediction Model Based On Feature Engineering

The law of users' purchase is generated by the interaction between the user's own internal and external factors.Therefore,the law of users' purchase not only has the certain specific rules to follow,but also has the certain complexity and uncertainty.With a lot of users' actual purchase data accumulated by the shopping platform,the laws of the users' purchase are discovered,and a reasonable way to predict is adopted to solve the problem of user' purchase prediction,which has the certain guiding significance for the personalized marketing of ecommerce.This paper has researched on user' purchase forecasts.The specific work is summarized as follows:Firstly,according to the actual users' purchase data provided,the existed data sets are visually analyzed and processed,and the users' purchase rules are found,and we constructed the reasonable data characteristics depending on the user-predictive data sets.Secondly,use the determined experimental data sets to study the users' purchase prediction model,select the random forest model,the GBDT model and the LightGBM model,and combine the acquired experimental data sets to design the user to purchase the single prediction model,and instantiate the results according to the single prediction model.The analysis determines the single model that the users buy the combined forecasting model.In the model training,according to the influence of different features on the prediction accuracy of the model,and the core idea of the tree model based the method and characteristics of the LightGBM model,the feature is selected.The final experimental data set,combined with the final experimental data set for single prediction model practice,the experimental results verify that the prediction accuracy of the model constructed after feature selection is better than the model constructed before feature selection.Finally,according to the results of the analysis of the prediction,the GBDT model and the lightGBM model are selected for model fusion to obtain the final user purchase prediction model,the single model in the fusion model is grid-adjusted,and the final experimental data set is used for training prediction.The experimental results show that the prediction accuracy of the GBDT-LightGBM which is a combined model is better than the single prediction models.