Neural Network Based Non-parametric Winning Price Estimation Of Display Advertising

In recent years,real-time bidding display advertising has become an important component of online advertising by virtue of its active push feature.Meanwhile,the emergence of real-time bidding display advertising have drawn huge attention of winning price estimation.Precise winning price estimation allows advertisers to design optimal bidding strategies,so as to win more valuable advertisement impressions with lower costs and gain more profits.Most of existing approaches use censored regression to model the historical winning price data based on the assumption of a single parametric distribution with survivor analysis.However,due to the complex distribution of winning price in actual scenes,this assumption is often too restricted,resulting in poor generalization ability of these models.Thus,in order to solve this issue,we propose a Non-Parametric Neural Network based Winning Price Estimator — NPNN-WPE.On one hand,NPNN-WPE forecasts winning price with non-parametric neural network,which can adjust network structure automatically during the training process.This model can complete the adaptation of training data and adjust the number of model parameters in one go,reduce the overall training time significantly and achieve better performance.On the other hand,NPNN-WPE adopts two non-parametric methods to process right censored historical winning price data caused by the second price auction of display advertising(only the advertiser with highest bid can observe the real winning price,when other advertisers can only know the real winning price is higher than their bids).The first one is weighted advertising bidding data based on the survivor analysis.The second one is the application of a specifically designed loss function,which is based on non-parametric censored least absolute deviations estimation.Methods based on non-parametric estimation can reduce training computational complexity and improve generalization ability without any functional assumption about winning price distributions.Experiment results on the i Pin You dataset and Criteo dataset widely used in the study demonstrate that the prediction performance of the model proposed in this paper is significantly improved compared with the parametric model proposed in previous studies.