The Study Of The Shopping Platform Advertising Recommendation Based On Conversion Rate

Thanks to the popularity of Internet technology,the e-commerce field is developing rapidly,more and more products can be purchased on the Internet,and the number of online shopping users is increasing.This status promotes the rapid development of online advertising.Among them,search advertising is the form of online advertising with the fastest growth,the largest scale,and the highest revenue.It is the interaction process between merchants(advertisers),users and e-commerce platforms.The merchants sum up the characteristics of their own goods,and then purchase the corresponding keywords.When a user searches for these keywords,the platform displays the corresponding ad on the user's search page.How to push a more appropriate ad to a user,ensuring that the user clicks on the ad and generates a purchase has always been a research focus in the field of search advertising.This paper studies the ad recommendation problem based on conversion rate,and constructs a parallel model combining traditional recommendation model(FM)and deep learning model(DNN and DCNN)to explore the data and improve the accuracy of conversion rate forecasts.After that,the estimated conversion rate is sorted,and then based on the Top-N strategy,the top 10 ads are selected and recommended to users.This paper first preprocesses the data of the real advertising data set of the Taobao E-commerce platform.Secondly,three single models of factorization machine model(FM),deep neural network model(DNN)and dynamic convolutional neural network model(DCNN)were established to predict the conversion rate.Compared with the traditional recommendation model,this paper adds additional features such as contextual features and product attribute text features processed by word embedding technology.These additional features can supplement more detailed user's and product's information from different angles.Thirdly,two evaluation indexes of log loss and AUC were selected to evaluate the effect of the model.In contrast,the results show that the dynamic convolutional neural network model works best.Aiming at the shortcomings of single model,a combination model using parallel structure is constructed,which combines the advantages of FM model in low-order feature combination and the advantages of DNN and DCNN model in higher-order feature combination,which further improves the prediction effect.The logarithmic loss of the combined model is 0.1382 and the AUC is 0.8079,both higher than the individual models.Finally,this paper summarizes the main work and proposes a follow-up improvement direction for the shortcomings.The research results of this paper show that based on the multi-dimensional characteristics of users,commodities,contexts,stores and text features of e-commerce platform advertising dataset,the combination model of deep learning technology and traditional recommendation technology can achieve better prediction results than single model.According to the results of the data test,the combined model can effectively identify the user's conversion behavior in the real environment,thereby improving the recommendation effect.