Research And Implementation Of Bidding Algorithm In Real-time Bidding System

With the rapid development and popularization of Internet technology,more and more advertising gradually changed from offline to online.Nowadays Real-Time Biding(RTB)is the most popular method of ads delivering in display advertising,which is a type of Internet advertising.In RTB,whenever a chance of advertising exposure arrives,the DSP will calculate the bidding price according to the rules.And the bidding price will directly determine whether this ad can be delivered,which becomes a key point in this problem.As a new way of advertising,there are large optimization spaces in the key algorithms of RTB,which this thesis aims to deal with,especially the bidding price optimization problem.This thesis firstly researches the bidding problem under limited budget,modeling it as a revenue optimization problem under limited budget constraint,and then explores problems of click through rate predicting and budget managing.In the aspect of click through rate prediction model,this thesis proposes a prediction model based on tensors and deep learning.This model utilizes the three aspects features of the ad to build a three order tensor by taking the direct product of them,then learns the hidden weights of every feature,and puts these weights into the deep learning model to further train its parameters.In the bidding and budget management aspect,this thesis designs a prediction model based on historical price and bidding failure rate,so as to effectively enhance the effectiveness of the demand side platform bidding algorithm.Besides,considering every ad has its own budget limit,in order to spend budget smoothly,here in this thesis we designs a pre-allocated budget and real-time adjustment scheme based on periods of advertising time and winning rates of bidding.On the iPinYou datasets we evaluated the performance of the proposed CTR prediction model,and compared with other state-of-art models.The results show that the predicted AUC index of the model proposed in this paper is better than the factorization machine up to 2.7%,better than the logistic regression prediction by 3.8%.Experiments on the iPinYou show that the biding algorithm based on winning rates prediction improved the number of clicks by 2.8%.And the budget management algorithm improved the number of click by 8%.