Winning Price Estimation For Real-Time Bidding Display Advertising Based On Soft Decision Tree

Real-Time Bidding(RTB)is an emerging and influential display advertising buying mechanism in the era of big data.Based on the analysis of a large amount of data generated by Internet users,RTB system can identify the characteristics and interests of the target audience of each impression and automatically serve the best matching advertisements.The most concerned research work in RTB system is related to Demand-Side Platform(DSP).DSP represents the appeal of advertisers.The quality of DSP bidding strategy directly affects the ability of advertisers to obtain high-quality traffic,thus affecting the conversions from advertising marketing.In DSP,the winning price,as the cost of the impressions,can effectively guide the formulation of the bidding strategy and the allocation of the budget.This paper has carried out related research work on Winning Price Estimation and proposes a Winning Price Estimation model based on soft decision tree,which solves two problems:(1)the current research work on Winning Price Estimation is mainly based on the assumption of winning price obeying a parametric form of function distribution.But in practice,the winning price comes from the bidding of hundreds of advertisers for an impression,which does not simply obey a hypothetical function distribution.Decision tree model can avoid this problem,Decision tree provides a selflearning process from input to output without functional hypothesis;(2)the construction process of the ordinary decision tree ignores the possible correlation between the dimensions,and the soft decision tree is different from the ordinary decision tree.Soft decision tree can influence the left and right child nodes according to different probabilities when the nodes are split.This feature can be used to model the correlation between dimensions.In addition,since the RTB process adopts the second-price auctions,the winning price can only be observed when the bidding is successful,and the failed bidding can only know the corresponding bid,so there will be a problem of censored in Winning Price Estimation.In this paper,the soft decision tree model is improved by combining survival analysis.The KaplanMeier Product-Limit method is used to estimate the probability of successful bidding without bias,and the Inverse Probability of Censoring Weighted method is used to correct the real error of the winning price,so that the data records of successful bidding and failed bidding can be feedback to the learning of the soft decision tree model at the same time,which reduces the Model deviation caused by censored data in practice.This paper designs and conducts experimental verification based on the real display advertising dataset which named iPinYou.The experiment compares the winning price Estimation model based on hypothesis distribution and ordinary decision tree respectively.The experimental results show that the proposed model performs better than other models in predicting errors.At the same time,the validity and necessity of considering censored data are verified.Finally,in the comparison of tree model scale,the proposed model has a smaller tree model complexity.