Research On User Purchase Behavior Prediction Based On Heterogeneous Integration Algorithm

With the rapid development of e-commerce and the popularization of Internet?Internet of things,the information of Internet users is increasing day by day.At this stage,we have entered a big data era with information explosion and"Big Data" has become one of the important symbols of modern society.The operation behavior of users to commodities in e-commerce big data can reflect users' preferences,how to mine users' potential preferences from massive real data has become the research focus of academic and industry,which has achieved many research results.At present,the research on the prediction of user's purchase behavior has not taken into full consideration the behavior characteristics of multiple types of users and the relationship between different behaviors,and the application of integration algorithm is mainly homologous integration.In this view,this thesis constructs the overall user behavior characteristics from five aspects on real data sets,considering the representational learning ability of deep learning and the training efficiency of integrated learning comprehensively,a multi-layer heterogeneous integration algorithm was introduced to combine Random Forest(RF),LightGBM,XGBoost and other algorithms,and a multi-layer algorithm framework based on deep forest and Stacking learning was established to predict users' repeat purchase behavior and purchase date.The main work of this thesis is as follows:1.In terms of basic preparation for user purchase forecast,this thesis introduces the research status and achievements at home and abroad in detail,and describes the algorithm models widely used in this field.Such as traditional machine learning algorithm:logistic regression,support vector machine and decision tree,homologous integration algorithm:random forest,XGBoost,LightGBM and convolution neural network,and introduces the theoretical basis and model architecture of the above model.2.In the aspect of time sliding window feature construction,a comprehensive study of the correlation theories and key technologies of feature extraction?feature selection and feature construction is conducted in this thesis.Based on the common sparsity and timing characteristics of user behavior data sets in e-commerce platform,this thesis innovatively introduces time sliding window technology to construct features under different time windows,and proposes a feature construction method of window weight decreasing.Finally,249 dimensional feature vectors are extracted as training data sets from five aspects:basic features,time series features and correlation features and so on.3.In the aspect of user purchase behavior prediction,this thesis studies from two aspects:First,based on the theoretical basis and technical optimization of deep forest algorithm,we proposes a method of user purchase behavior prediction based on multi-Grained Cascade forest,which uses multi granularity scanning module to transform 249 dimensional input features into 1800 dimensional instance features,and uses deep cascaded forest to represent the learning results layer by layer.Then,it proposes a method of user purchase date prediction based on Stacking,which is used to further predict user purchase date.The experimental results show that the above algorithms have good prediction accuracy and training time.Thus,it is proved that the application of heterogeneous integration algorithm in the field of e-commerce is effective.