Research On Personalized Purchase Prediction Based On User-Product Demand Driven

The timely acquisition of user feedback information on products is very important for product improvement design,manufacturing,logistics and transportation and sales strategy based on user needs.Data mining technology can be used to analyze the user-product interaction behavior logs stored in the online trading platform,explored user preferences and purchase intention,and predicted user's purchase behavior.Therefore,the feedback information of users on some aspects of the product can be analyzed as early as possible,which can guide the further improvement of the product,adjust the production plan,manufacturing technology and transportation arrangement,and provide important reference information for the sales enterprises to formulate appropriate marketing strategies.The prediction of user's purchase behavior can be regarded as a supervised binary classification problem.This paper analyzes user-product interaction data through data mining technology,establishes feature engineering to predict user purchase behavior,and proposes a weighted integrated model based on stacking to predict user purchase behavior.The main research contents are as follows:(1)The data of user product interaction and other data acquired are preprocessed.Preprocessing includes data description,data collation,data cleaning(missing value processing and outlier filtering)and exploratory statistical analysis of user product interaction data,which can deeply mine user behavior characteristics and overall data distribution,and provide effective reference information for the next step to construct the features of predicting user's purchase behavior.(2)Feature engineering is established for preprocessed user-product interaction data.The specific process of feature engineering for user purchasing behavior prediction is as follows: feature construction of user purchasing behavior prediction?feature scaling?feature coding?mixed sampling of positive and negative samples of feature set?feature discretization?feature selection,so as to obtain a balanced and effective user purchasing behavior prediction feature set that can be directly input into the model.(3)A weighted ensemble classification model based on Stacking is proposed to predict the final purchase behavior of users.The basic learners in the Stacking ensemble framework are set as Support Vector Machine(SVM),C4.5 decision tree,Logistic Regression(LR)model,and meta learners are LPBoost algorithm with soft intervals.The goal programming problem of LPBoost algorithm is solved to obtain the weight vector of the base learner,and a weighted ensemble classification model based on Stacking is established to predict the user's purchase behavior.The model proposed in this paper is compared with SVM,C4.5 decision tree,LR and majority voting method of these three basic learners to evaluate the performance of the models.The user-product data set officially released by Aliyun Tianchi is taken as the research object,and the experiment is carried out in combination with the above research contents.The experimental results show that the proposed scheme is better than the comparison scheme,which can achieve a high prediction accuracy.