Estimation Of Display Advertising Click-through Rate Based On USFD Method

With the rapid development of Internet and communication technology,the marketing model of "Internet+Advertising"(Online Advertising)has become one of the main channels for enterprises to promote goods and gain profits in the age of big data.Recently,display advertising which has been used widely and is the most important component of online advertising has grown strongly.Under the CPC(Cost per Click)billing model,whether the advertising recommendation system can accurately estimate the click-through rate(CTR)of display ads is the core issue in the research of industry and academia,because improving the accuracy of ads click-through rate can not only improve the income of advertising agencies and advertisers,but also enhances the user's experience.In this paper,the historical click logs of Taobao's display advertising are used as the experimental data set and the goal is to improve the accuracy of CTR's estimation.After the analysis of the experimental results about different characteristics in the single models of Logistic Regression Model(LR)and Factorization Machine Model(FM),and considering that the same characteristic has different influences to users,behavior on different user groups,we propose to apply the method of USFD(the Advertising Click-through Rate Estimation Based on User Similarity and Feature Differentiation)to improve the accuracy of the estimation of CTR.It mainly includes the following aspects:Firstly,this paper has introduced the meaning of the experimental data set and each character of the data set in detail.Then we has extracted the training set and test set and processed the continuous feature and discrete features by segmenting and one-hot coding respectively.Finally,the ROC curve and AUC score are used for evaluating the model performance and the reasons for choosing them are given.Secondly,a single LR model and a single FM model are used to predict the click-through rate of ads.By taking the specific Features,classification Features and all features as the input variables of the models,it has found that the FM model with all features has the best estimation results,which lays a foundation for the use of USFD method when we select the prediction sub-model.Finally,because the single prediction model does not deal with the relations of user and user characteristics,so the USFD method is used as combined model to estimate ads CTR.It first uses the Gaussian Mixture Model(GMM)to cluster the users according to their similarity,and then uses the FM Model to estimate the CTR based on all features of the data set in each user group.For each test sample,the similarity with each gruop is taken as the weight of the CTR estimated by the corresponding group's sub-model which the final prediction result of the test sample is obtained.The experimental results show that the USFD method is better than the single model,and its AUC score is 19.1%and 9.6%higher than the LR model's and the FM model's respectively.