Predict Users' Purchase Intention And Target Product Category Based On The Characteristics Of User's Behaviors

Today,with the rapid development of e-commerce,active marketing can make merchants stand out from the market flooded with commodity homogenization,attract users,and effectively improve the conversion rate of user consumption of marketing.Traditional methods for active marketing include advertising and media,which rely on traffic and luck to gain the conversion rate.Therefore,in order to find a more effective way to improve the conversion rate of user consumption,the key lies in how to accurately acquire the target users and provide them with the information on the products they are most likely to buy.How to acquire the target users and products involves the issue of mining and prediction,which is right the research content of this paper.Under the tide of mobile Internet,the behavior of users in the network environment occupies an ever-growing proportion.Users' clicking,browsing,purchasing,commenting and other behaviors on the Internet have become very important reference data for the service providers.To realize the accurate prediction of user behavior,this paper takes the behavior data of e-commerce as the entry point,analyzes and models the data,and predicts the customers' purchase intention and the product category they are most likely to buy in the following week with different machine learning methods,which will help the merchants pinpoint the user group with the highest conversion rate,develop more precise marketing strategies,and further improve the promotion of sales conversion rate.The specific modeling steps are as follows:The first step is the data acquisition and preprocessing.Collect the basic information and behavior data of the users logging in the website and the commodity data of e-commerce,convert the relevant data into the format required by the proposed method,and perform abnormal data cleaning.The second step is the data feature extraction.By encoding the basic data,the basic features of users and commodities are extracted.The basic user behavior data are counted from different dimensions to generate statistical characteristics,time interval characteristics,and computational characteristics.The characteristics are associated and fused based on the relationship between them.The behaviors are classified and weighted using the theory of time decay.Then,Chi-square filtering is used to calculate the chi-square statistics between each non-negative feature and label,and the K features with the highest scores will be selected.The third step is the processing of positive and negative samples.According to the users' actual purchasing data,mark the user characteristic data,generate positive and negative samples,automatically analyze the imbalance of them,perform downsampling according to the preset proportion,and generate a subset of positive and negative samples.The fourth step is the training and prediction of the model.RF,GBDT,and XGBoost algorithms are used to train the model with the samples and generate multiple prediction results for comparison.In the fifth step,the limitation of the model is analyzed by comparing the prediction results on a single undersampled sample of different models.In order to solve the balance of positive and negative samples,and make full use of the numerical superiority of the samples,a novel downsampling method is proposed to extract the subset of positive and negative samples for several times,which is adopted by the RF algorithm as the random sampling way to customize RF training model.Finally,the rationality and improvement direction of the model are evaluated by comparing and analyzing the prediction results of the single undersampled sample of different algorithms in step 4 and the prediction results of the multiple training models with different sample subsets in step 5.