Prediction Model Of Advertising Real Time Price Based On Thompson Sampling And Censored Regression

In Real Time Bidding(RTB)system,the Demand Side Platform(DSP),which represents the interests of advertisers,needs to maximize the return on investment(ROI)of advertisers.Therefore,it is important to accurately predict the transaction price of the advertisement to improve the ROI.In general,the DSP can establish a linear regression advertising price prediction model by using the historical bid records data they hold,there are two problems.First,under the current RTB rules,the DSP can only know the advertising price while the auction succeeded,which is the price they should pay,if the auction failed,they can only know the lower bound price of advertising,that is,their own bid.So the advertising price in the historical auction data held by the DSP is censored data.In this case,the linear regression model will import a large deviation whether it removes the censored samples or uses all the samples.Second,as time goes by,advertising and publisher will present the trend of rising or declining,the user's interest will change too.Therefore,the distribution of the advertising price will change over time,resulting in the instability of the prediction model.Aiming at this background,this paper proposed a mixed censored regression model to solve the problem of data censored and improved the accuracy of single prediction model.And then,with training multiple regression models and using the Thompson sampling algorithm to select and fusion,to a certain degree,improved the accuracy and stability of advertising transaction price prediction.First,this paper investigated the background of RTB market environment,bidding mechanism,payment method and common optimization goal of DSP,and then studied the development status of advertising price prediction and the tobit regression model applicable to censored data.On the issue of model selection,the problem of Multi-Armed Bandit(MAB)and the extensive application of Thompson sampling algorithm are studied.Second,the hierarchical label system for advertisers,users and publisher was built on existing advertising bid data.And then,the data was split into several subsets and trained into several models according to different level label combinations.Then,by using the method of Thompson sampling,the appropriate model is selected to predict the advertising price according to the principle of random probability matching.Finally,the assumptions and methods described above,which was based on iPinYou and bridgewell's open auction advertising datasets,were verified first.Then,a series of experiments were carried out on the proposed algorithm and some baseline methods.The experimental results,which were fully analyzed,proved the validity of the model and the model we proposed has certain practical application significance.