Research On Precision Delivery Of Consumer Vouchers Based On XGBoost And CatBoost

O2O,Online to Offline,is a business model in which offline merchants and the Internet are combined.In recent years,with the rise of mobile network consumption,O2O consumption has received extensive attention from major commercial platforms,which has great commercial value.Consumer voucher placement is an import marketing strategy of O2O,however,random delivery will interfere with most users and increase the marketing cost of merchants.Precision delivery is an important technology to improve the write-off rate of consumer vouchers,it enables merchants to directly face users with certain consumer preferences,so that they can get real benefits,while reducing the marketing costs of merchants,promote business turnover and achieve win-win results for both merchants and users.The O2O industry's characteristics make it naturally associated with hundreds of millions of consumers,recording a huge amount of user history behavior.This paper excavates these real user history behavior records and constructs the corresponding feature group training to get the user's consumption voucher usage prediction model,so as to predict whether the user will use the consumption voucher in the future,providing decision-making basis for the accurate delivery of the consumption voucher.The main research work of this paper can be divided into four aspects:(1)Firstly,we explore the data from three aspects: consumer vouchers,merchants and users,and then construct corresponding feature groups to represent users' consumption behavior habits and information attributes of merchants and consumer vouchers.Then,we filter the original features through variance selection and extreme gradient boosting tree algorithm XGBoost to remove the corresponding redundant features and complete the construction of data sets;(2)In order to enhance the representation ability of features,this paper transforms the original feature group based on XGBoost and merges the transformed features with the original features to form a new training data set,and then trains XGBoost to get a consumer voucher usage prediction model based on XGBoost feature transformation.The experiment shows that compared with the original feature set,feature transform can further improve the prediction effect of the model;(3)Traditional gradient boosting algorithms generally produce biased gradient estimation in the training process,which results in prediction shift of the final training model and affects its generalization ability.For this reason,this paper introduces a new gradient boosting algorithm CatBoost,which is based on Ordered Boosting to train,and obtains unbiased gradient estimation for training to slow down prediction shift,enhancing the generalization ability of the model.The experiment shows that the CatBoost-based consumer voucher usage prediction model has better model performance than the traditional algorithm;(4)Finally,the single models in(2)and(3)above are weighted and fused to obtain the final consumption voucher usage prediction model based on XGBoost and CatBoost,which can predict whether users will use vouchers in the future.The experiment shows that the fusion model has better prediction effect than the single model,which can provide a more reliable decision-making basis for accurate delivery of consumer vouchers.