Application Research Based On The Integrated Learning Model In The Forecast Of The Number Of Shoppers

With the applying of Internet technology and the diversification of payment methods for consumers,it is particularly important to monitor data for online retail.At the same time,APP has accumulated a large amount of users and consumption data,and more and more attention has been paid to how to transform huge amounts of data to be informations with commercial value.Predicting the number of consumers about stores can optimize operations and reduce costs.Therefore,there is an urgent need for forecasting the number of consumers about stores.Based on researchs at home and abroad,this paper collected stores' information through the software on its own operating platform.Through simple time series model and ensemble learning method(random forest,GBDT,XGBoost)to predict the number of consumers in the next period and evaluate the effect of models.The empirical part uses nearly 600,000 customers' data of actual business,and establishes scientific and rational feature engineering framework by understanding the model principle and the actual business background.Besides,through the framework exploring and visualizing data.The resulting feature dimension is 275 dimensions,and the training sample length is 469,575.By using these features to establish the store consumption forecast models and by comparing loss function and mean squared error about the simple time series prediction model and the ensemble learning methods(random forest,GBDT,XGBoost)to give the advantages and disadvantages.The final result shows that the ensemble learning method XGBoost is superior to GDBT in prediction accuracy,random forests,and simple time series prediction models,and have a good advantage in overfitting problems,but the running time is in the middle.Overall,XGBoost is excellent.Finally,summarizing the full text from the importance of feature engineering,the adjustment of parameters and the accuracy of the final model prediction,and proposes follow-up research suggestions for the inadequacies of this paper.