Research On Prediction Of Customer Purchase Amount Based On Heterogeneous Fusion Model

With the rapid development of Internet technology and the continuous improvement of network infrastructure,more than 780 million people in my country have chosen to shop online,and a large number of users’ purchasing behaviors have generated massive data.Learning technology to predict the future purchase behavior of customers has become a research hotspot.Purchase amount forecasting refers to forecasting the total amount of orders purchased by customers in the future.Enterprises can use purchase forecasts to adjust stocking strategies,formulate sales targets for business personnel,and formulate operating budgets for the next cycle.In order to predict the purchase amount of users,this paper takes the consumption data of online wholesale e-commerce as the starting point,conducts in-depth analysis,mining and modeling of the data,and establishes different models to predict the purchase amount of customers in the next month.The specific work includes: :(1)Data exploration and feature screening: First,analyze the problems to be solved in this paper,sort out the relevant factors that may affect users’ purchasing behavior,and explore and analyze the target amount through data visualization.Then,the data is preprocessed,and features are reconstructed according to the data analysis results and business experience,and three feature groups are constructed,namely time series feature group,business feature group and attribute feature group.Finally,in order to reduce redundant features and improve model training efficiency and accuracy,the effects of various feature screening methods were compared on the external evaluator.The experimental results show that the feature screening method based on SHAP(SHapley Additive ex Planation)has the best effect.(2)Prediction model design: First of all,in the traditional Stacking algorithm,the average value is used to form the test set in the K-fold cross-validation for optimization and improvement,and the strong features of the original training set screened based on SHAP value are added to the meta-learner.In order to improve the predictive ability and robustness of the model.Then,four heterogeneous(that is,different principles)single models are selected as the basic learners of the Stacking model,and the purchase amount prediction models are constructed based on the traditional Stacking algorithm and the improved Stacking algorithm respectively.(3)Single model prediction: Models based on Random Forest,Light GBM,LSTM,and Tab Net were constructed to predict customer purchase amounts,and Bayesian Optimization was introduced in the model training phase to adjust the hyperparameters of the model.excellent.The experimental results show that the prediction accuracy of each single model is better,but the prediction performance for different sizes of target quantities is different.(4)Heterogeneous fusion model prediction: First,four single models with better performance after hyperparameter tuning are used as the base model,XGBoost is used as the meta model,and the stacking model fusion method is used for fusion.The experimental results show that the improved Stacking model has the best prediction effect;then the fitting effect of each model is analyzed through data visualization;finally,the rationality of the model design is studied through an ablation experiment.