Research On Click-Through-Rate Prediction Algorithm Based On Deep Learning

The way people learn,live and work has been changed by the rapid development of Internet technology.Click-Through-Rate prediction is one of the core technologies in Internet search,recommendation and advertising.In the past few decades,the Internet advertising industry has developed rapidly.Advertising Click-Through-Rate prediction technology has become an important research content in Internet advertising.In recent years,with the rapid development of deep learning,the accuracy of ad's Click-Through-Rate prediction has been greatly improved.The accuracy of the ad's Click-Through-Rate prediction is directly related to the quality of ad's delivery,affects the advertiser's marketing performance and the revenue of the ad delivery platform,and also affects the user experience.Therefore,the prediction of ad's Click-Through-Rate is very important.At present,traditional machine learning and deep learning are mainly used in ad's Click-Through-Rate prediction.In order to further improve the accuracy of ad's Click-Through-Rate prediction,the optimization of algorithm and feature vector extraction has been adopted by academia and industry.However,the data in the scenario of Click-Through-Rate prediction is complex,sparse,and huge.Complex rules in data are difficult to fully learn by algorithms.It is also a major difficulty to represent the high-dimensional,sparse original feature vectors as low-dimensional,dense and easyto-learn vectors without losing information.Therefore,in order to improve the accuracy of ad's Click-Through-Rate prediction,two aspects of algorithm and feature vector extraction are proposed.In order to further improve the accuracy of the ad's Click-Through-Rate prediction,the main research contents of this paper are as follows.Firstly,in terms of algorithm,deep learning has great advantages in expression ability,Click-Through-Rate prediction algorithm is studied based on deep learning.Specifically,the classical Wide&Deep model is chosen for the study,the advantages and disadvantages of the Wide&Deep model are analyzed,in the deep part of the algorithm,the feature combination layer in the feature domain and the feature combination layer between the feature domains are added to make it more specific to learn the cross-over characteristics in advertising data,and to improve the generalization of the algorithm.Then,in the aspect of feature vector extraction,the method of extracting user interest features is studied,and two user interest representation methods are proposed,which are user interest representation based on interest clustering and user interest representation based on time clustering.The former clusters the original user interest vectors,and uses different methods to integrate the user interest vectors into a fixed-length dense vector within and between clusters,so that it can accurately represent the user's complex interests in specific occasions.The latter clusters the original user interest vectors in the time dimension,and takes the influence of time on user interest into consideration when integrating the user interest vectors into a fixed-length dense vector,and attenuates the early interest.Finally,based on the current research status of the industry and the existing experimental environment,the dataset is processed with the technology of feature engineering,and the experiment is designed on the advertising dataset,and the experimental results are analyzed.In terms of algorithm,in order to verify the validity of the proposed algorithm in ad's ClickThrough-Rate prediction,an improved Click-Through-Rate prediction method based on Wide&Deep model is implemented.Several classical algorithms including Wide&Deep model are compared.The results show that the algorithm proposed in this paper improves the AUC and Logloss indicators by 1.16% and 2.27%,respectively,compared with the Wide&Deep model before the improvement.in the aspect of feature vector extraction,user interest is represented by user interest representation method based on interest clustering and user interest representation method based on time clustering.The obtained interest vector is input into the improved Wide&Deep model.Compared with the traditional interest representation method,the accuracy of Click-Through-Rate prediction method is improved more significantly by the two methods presented in this paper.