Research On The Optimization Of Bidding In Real-Time Bidding Advertising

With the expansion of the number of Internet users,the operation mode of advertising campaign is also changing,it extends from the traditional offline mode to online mode.At present,one of the most important way to display ads is through realtime bidding system.In real-time bidding,demand side platform which is on behalf of the interests of advertisers evaluate the quality of impressions according to the factors,such as the target rules,the features of advertising campaigns and the impressions.And then combined with the budget amount,the conversion value,the DSP decide whether to participate in bidding,and offer the specific amount of bid price.The final output of the demand side platform is the bid price for impression,so the relevant algorithms on the DSP can be considered as part of the bidding strategy.As an indispensable component of real-time bidding system,the research and optimization of bidding strategy on the DSP is a noteworthy direction.This thesis divides the bidding strategy of into the following parts: click-through rate prediction,bidding landscape,bidding algorithm and budget management.This thesis will mainly focus on the last three parts.The innovative results of this thesis are as follows:(1)Considering the incomplete data information of the demand side platform,this thesis will put forward the bidding landscape model to consider the censored data.The model consists of two parts: the winning rate prediction algorithm and the winning price prediction algorithm.In this thesis,we utilize the product-based neural network model to predict the winning rate.Winning price prediction algorithm include three algorithm: we will use the FM model to model the winning and deleting data respectively.And then we will propose a mixture model,which combines the two algorithm,weighted by the results of the winning rate prediction model to combine the two models.(2)In order to win impressions that consistent with the delivery strategy of advertising campaign.This thesis will propose bidding algorithm which is based on the bidding landscape.The algorithm is also considering the quality of impressions.At the same time,in order to adapt to the dynamic changing bidding environment,this thesis also adopts the feedback mechanism to adjust the bidding function in real time.(3)Because the budget of each advertising campaign is capped,the nature of the bidding problem is the optimization problem of budget constraint.Therefore,budget management algorithm of this thesis will include budget allocation and threshold setting.Among them,the budget allocation algorithm is based on the traffic and the quality.The threshold setting algorithm sets the threshold dynamically according to budget of each the time slot,and will effectively filters the low-quality impressions,so that the budget will spent more on effective impressions.This thesis conducts the experiment on iPinYou dataset,the AUC of the winning rate prediction model which based on PNN reached 93%,and the MSE of the winning price prediction algorithm was 1151.Under two budgets,the proposed bidding algorithm improved by 4.9% and 2.7%,respectively,compared with the existing algorithms.And the budget management algorithm improved by 0.7% and 1.2% respectively.