Optimization Of DSP Advertisement Sorting Based On Visible Features

With the vigorous development of the advertising market,advertisers have invested more and more in network marketing in recent years.However,during the process of Display Advertising,not all advertising exposure can be seen by users.Accordingly,advertisers try to focus on distinguishing visibility of display advertising for saving advertising costs.Therefore,whether the advertisement is visible and corresponding impact have become a direction of technology research.DSP,Demand Side Platform,in short,it is a platform that allows different advertisers to advertise by looking for high-quality media,targeting customer groups,and optimizing their delivery strategies continuously.Such platforms can bid on Display Advertising based on user access information,platform operating data,and bidding algorithms.In this process,which ads are selected for display bidding need the ads been sorted in the advertising library of their platform.This article estimates that the visibility of ads has a positive effect on the ranking model of ads(involving the click behavior of ads and the conversion behavior of ads).During our research,firstly,we solve how to measure the AD visibility.Aiming at how to judge whether the display advertisement is visible,we use the auxiliary detection and construction of advertising monitoring point detection methods during the data collection process of advertising delivery.Secondly,after the data collection of advertising,the visible features are expanded based on both two method as statistics and model.On the basis,DEEPFM model is introduced to predict the CTR,DSSM model to predict the CVR,and ESMM model is used to train the CTR model and CVR model jointly.Finally,achieve the optimization expectation of advertising ranking of DSP by increasing CTR and CVR result after introducing visible feature.This thesis has the following core content:First,on the DSP side,it analyzes the visibility of an advertisement in the case of different sizes of advertising and browsers from the client side.It designs and collects related advertising visible data,constructs the visible characteristics of advertising.Second,based on the area of the advertising display,the thesis constructs the visible characteristics of the image according to the intention of the display part of the Ad image.By using the image truncation factor,it can more effectively predict the user's real intention of the display advertising of image type,and reduce the possibility of recommending irrelevant advertising to users.Third,features extracted manually is added to optimize the model of advertising ranking.The accuracy of advertising ranking model is improved after adding AD visible feature,and the advertising retrieval system of DSP is further optimized.The optimized DSP is tested via multi-dimension experiment based on real data set.The experimental results show that after introducing visible features,the accuracy of CTR and CVR is significantly improved,e.g.CTR is improved by 5.4% under DEEPFM,overall CTR is improved by 5% synthetically,CVR model after training jointly with visible feature is improved by 8%;at the same time,the introduction of image truncation factor helps to improve the effect of advertising CTR prediction,and we analyzed how to determine the optimal value of image truncation factor.So the accuracy of recommendation of advertising retrieval of DSP is significantly improved by 6.2% than original sorting system,consequently proving the effect of using visible feature.