Research On User Purchase Behavior Prediction Based On LightGBM

The concept of e-commerce is becoming more and more popular.Various e-commerce platforms are emerging one after another.More and more people join the army of online shopping.When the e-commerce platform develops to a certain extent,the increase of traffic will eventually stop.Improving the traffic conversion rate is undoubtedly an important and urgent topic.At present,all e-commerce platforms have introduced recommendation algorithms to recommend their favorite products for users and improve user experience.Prediction is the basis of recommendation.Predicting the purchase tendency of users in advance will undoubtedly greatly improve the effect of recommendation algorithm,which is a very meaningful work.Based on this,this paper selects the data of Jingdong algorithm contest to study the prediction of users'purchasing behavior.The main work and results are as follows:1.Determine the prediction target:in the user category store combination(called F₁ID)with behavior records in a period of time,it is predicted that the F₁ID of purchase behavior will be generated in the next seven days,which is a binary classification problem in prediction.2.Determine the samples of training set and prediction set.A total of 517049F₁IDS with behavior records from March 19,2018 to April 1,2018 were selected as training samples after negative sampling with a positive and negative sample ratio of1:30,in which the tag with purchase F₁ID in the next 7 days was 1,and the rest was 0.A total of 1792209 F₁IDS with behavior records from March 26,2018 to April 8,2018 were selected as prediction samples,in which the tag with purchase F₁ID in the next 7 days was 1,and the rest was 0.3.Construct features based on time sliding window.This paper constructs 564dimension features from three aspects of basic features,cumulative features and time sliding window features in five dimensions of user,category,store,user category and user category store,and deals with the missing values of features.4.Build the model and select the final prediction model.In this paper,LR,RF,GBDT,XGBoost and LightGBM are used to train the model on the 517049*565training set,and to predict on the 1792209*565 prediction set.From the AUC,F₁score,training time and other aspects,LightGBM is selected as the final model.