Prediction Of User Purchase Behavior Based On Machine Learning

With the rapid development of the Internet,the scale of consumers and commodities of E-commerce is constantly expanding.E-commerce not only provides convenience for users to shop,but also brings the problem of information overload.Based on purchase prediction,personalized product recommendation and marketing programs can be designed to help alleviate information overload and optimize the purchasing experience of users.The preferences of users are hidden behind the massive user behavior data in the E-commerce platform.Therefore,based on this kind of implicit feedback data,this thesis proposes a purchase prediction model to predict the purchase behavior of users.The main work and innovation points of this thesis are reflected in the following aspects.In terms of feature engineering,the feature dimension of implicit feedback data is limited,and most previous studies are based on a single time window.Based on the real interaction records of users and items in E-commerce scenarios,this thesis constructs features by using the time sliding window method.In order to explore the preferences of users in different periods,this thesis sets three different length time windows: longterm,medium-term and short-term.The features are constructed from the five dimensions of user,item,category,user-item and user-category.Finally,a total of 84 features are constructed,which effectively expands the feature dimension of implicit feedback data and provides ideas for feature construction.In terms of model construction,this thesis uses three cutting-edge machine learning algorithms of XGBoost,LightGBM and Cat Boost to predict user purchase behavior.The results show that the LightGBM model has the highest F1 score and better performance on imbalanced data.In order to further improve the prediction effect,this thesis constructs a homogeneous fusion model based on parametric perturbation.The model uses the Bagging method to fuse several LightGBM models.In addition,parameter perturbation is introduced into the base learners,which further increases the difference between the base learners.The experimental results show that the F1 score of the homogeneous fusion model based on parameter perturbation reaches 0.6107,which is better than the single model.