Research On Click Through Rate Prediction And Dynamic Bidding Strategy In Display Advertising

Recently,network traffic has risen dramatically and more and more publishers take online advertising as an important way to make profit.The research of computational advertising mainly focuses on finding the best match among a given user,a given context,and a suitable advertisement.The precise delivery of advertising benefits users,publishers and advertisers.As a new paradigm of display advertising,real-time bidding(RTB)not only changes the pattern of the advertising market by transforming the selling of ad slot from offline to online,but also greatly expands the research field of computational advertising.Therefore,the study of RTB algorithms becomes a hot topic both in the academia and the industry.In this paper,we divide RTB algorithms into two key issues:click-through rate(CTR)prediction and the design of bidding strategy.On one hand,CTR affects the benifits of the publishers,advertisers and users as mentioned before,on the other hand CTR is an important reference for making bidding strategies.However,the impression logs has serious data sparseness and it is difficult to achieve high prediction accuracy for traditional machine learning models.Due to the fact that advertising is an user oriented commercial activities,in this paper,we propose a CTR prediction model based on user similarity and feature differentiation(USFD)by considering that features have different effects on different user groups.In this model,we use clustering algorithm to divide users into different groups according to history logs and then train sub-classification model for each group.For a new trim(user,page,advertisement),we first evaluate the similarities between user and groups and then calculate the probabilities under the different classification models.Eventually,we determine the user’s CTR by a weighted combination of similarities and probabilities.According to the mechanism of RTB,advertisers obtain impressions through the auction raised by Ad exchange(Adx).Constrainted by the advertiser’s budget,a reasonable bidding strategy affects the advertiser’s ROI directly.Too high bid per auction will result in rapid consumption of budget while low bid will have little chance of impression.The current mainstream strategies mainly focus on static strategies or continuous feedback models.Considering the complexity of the Internet,we raise a dynamic bidding strategy based on probabilistic feedback(PFDBS)on the basis of CTR.In the strategy,we introduce deviation rate to evaluate validity of current strategy,furthermore,we give an approach to amend strategy combined with previous feedback when necessary.Finally,we practice several experiments on real world RTB dataset and make detailed comparative analysis with mainstearm approaches.Through analysis on Logloss and PR curve we can conclude that our CTR prediction model performs better among chosen menthods.Besides the best improvement of relative AUC is about 5%.In addition,our model can exploit the different influences on different groups for each feature.In the bidding experiments,by comparing the KPI and consumption on different strategies,we can find that the PFDBS can improve much ROI(nearly three times)of each advertiser under limited budget.Last but not least,the accumulated consumption of our strategy makes smallest error with that in actual market and keep the same trend of consumption with it.