Research On Advertising CTR Prediction Method Based On Deep Learning

With the rapid development of Internet,Online advertising has developed rapidly due to its advantages such as wide coverage,strong pertinence and high flexibility.How to maximize the number of clicks and conversions of advertising has become an urgent need for precision marketing.However,the existing click-through rate prediction models usually adopt a simple feature interaction method and fail to pay attention to the importance of features.At the same time,the existing models do not effectively mine the historical interest of users.In response to the above problems,this paper has carried out research from two perspectives of feature learning and user historical interest mining.The main work of this paper includes:(1)Aiming at the feature interaction problem of the existing click-through rate prediction model,the SeDeepFM model is proposed.This model based on the DeepFM model,by increasing the non-linear interaction of features so that the new model can dig out the information of second-order feature interactions more effectively and learn the importance of features by introducing SENet network.Experiments show that the new model can effectively improve the accuracy of click-through rate prediction。(2)Aiming at the fact that the existing click-through rate prediction models fail to dig out the historical interest of users accurately and fail to pay attention to the influence of historical interest on whether the target advertisement has been clicked or not,the SACSN model is proposed.The new model improves the transformer network to learn the historical interest of users and the correlation between the historical interest of users and the target advertisement.In order to further improve the accuracy of prediction,the auxiliary loss function is used to supervise the improved transformer network.Experiments show that the new model can effectively improve the accuracy of prediction.(3)Aiming at the fact that the existing click-through rate prediction models fail to mine the evolution information of users’ historical interest effectively,the IESACSN model is proposed.On the basis of the SACSN model,by increasing the interest evolution layer,user’s interest evolution information can be further captured while extracting user’s historical interest.Static feature interactions and user’s dynamic interest modeling are combined at the same time.Experiments show that the new model enhances the estimated effect of ad click-through rate models.To sum up,in order to solve the problem of low accuracy of the existing clickthrough rate prediction models,this article proposes three click-through rate prediction methods from the two aspects of feature interactions and the effective extraction of user historical interest: based on improved DeepFM method and based on user interest evolution method.And three corresponding models are proposed: SeDeepFM,SACSN,IESACSN.Experiments show that these three models can effectively improve the accuracy of click-through rate prediction.