| Recent years,under the role of market demand,display advertising has brought great opportunities for the development of the online advertising industry due to its huge network traffic,which has become a major source of revenue for e-commerce platforms,video sites,social media and other Internet enterprises.Real-time Bidding(RTB),an emerging display advertising trading model,has been widely used in various business environments,which consists of Double Click,Facebook Exchange and TANX.Not only does RTB itself give the advertisers opportunities to participate in the sale of advertising resources via auctions,but also its auction cycle being less than 100 milliseconds lets the ad audience have a better user experience.Unlike the traditional guaranteed delivery,it is by allowing advertisers to implement diverse display strategies for different traffic and audiences that RTB brings disruptive changes in advertising delivery and transaction patterns.Therefore,since the main trading platform declared its support for RTB in2009,RTB mechanism has become a popular and important issue in the research of academia and industry.Based on the above background,this dissertation focuses on the optimization of bidding strategy and the prediction of market price distribution in RTB system.The main research work is as follows:1)In RTB system,to maximize the key performance indicators for advertisers(such as number of clicks,number of conversions,ad revenue,etc.),demand-side platforms(DSPs)with large-scale data processing capabilities need to develop efficient bidding strategies.To achieve the goal of raising revenue,many existing bidding strategies tend to bid for bid requests with higher conversion rate.However,the size of ad audience with high conversion rate is limited in real scenarios.Moreover,due to the high commercial value of these audiences,it is common to attract various advertisers to bid at high prices,which leads to insufficient ad cost,and ultimately can not achieve higher ad revenue.Aiming to solve this,this dissertation theoretically analyzes the complex relationship between conversion rate,ROI and ad cost,as well as the dynamic relationship between these three factors and ad revenue.On this basis,this dissertation puts forward the Maximizing Revenue(MR)bidding strategy under certain ROI conditions.Unlike the existing methods,by defining the winning function as a tanh form,MR provides a more effective solution for accurately estimating ad cost and thus increasing ad revenue through maximizing ad cost under the condition of satisfying certain ROI constraints.2)In RTB system,accurately predicting the market price distribution is the basis for developing an efficient bidding strategy.Whereas,only the highest bidder can observe the true market price of the auction in the second-price auction mechanism widely used by RTB.Therefore,the data is right-censored,leading to the bottleneck in accuracy and generalization ability of traditional methods of predicting the market price distribution.In recent years,many studies have used Kaplan-Meier estimator in survival analysis to solve the problem of right-censored data.However,in these approaches,all samples in a population sharing a same prediction makes it impossible to discern a unique pattern in a single ad display.By analyzing the open datasets,this dissertation finds that features of ads have an important influence on the market price distribution.Based on this,this dissertation proposes the KMMN model that combines Kaplan-Meier estimator with Markov network to predict the market price at single ad exposure granularity by taking full advantage of different ad features,which reflects the influence of high frequency dynamic market price on bidding strategy more objectively.3)In order to predict the market price distribution,most of the existing methods are based on the assumption of unimodal distribution and ignore the obvious multimodal characteristics of the market price distribution in RTB system,thus seriously reducing the accuracy of the prediction.A Gaussian Mixture Model(GMM)is proposed based on research work 2 so as to solve this problem.GMM calculates the parameters of multiple Gaussian distributions and their mixing ratios based on the ad features,which fully reflects the multimodal characteristics of market price distribution under RTB scenario.On this basis,in order to solve the right-censored data,this dissertation extends GMM to CGMM,which improves the robustness of the model.4)Since 2019,the RTB system gradually changed from second-price to first-price auction mechanism.In first-price auction,the winning DSP pays its bid price for ad cost.For effectively reducing the winning premium(i.e.,to win the auction at the lowest bid price),DSP needs to accurately predict the market price distribution.In addition to facing the right-censored data and the multimodal characteristics of the market price distribution described in research work2 and 3,the left-censored data also needs to be addressed,since the winning DSP in first-price auction can only know that the market price is less than its bid price.Nevertheless,solutions to such problems are rare in existing research.On the basis of all the above background,this dissertation proposes the DMDN model,which uses a deep mixture density network to model the multimodal characteristics of the market price.By introducing a winning indicator,DMDN calculates corresponding probability of winning and losing bids and then designs the loss function for the first-price auction and second-price auction respectively,thus making it better to deal with the multimodal characteristics of the market price and the left/right-censored data under both auction mechanisms,and finally generates more accurate predictions of the market price distribution.All of these methods have been extensively compared across multiple large-scale open datasets.The results show that these methods are significantly superior to the advanced models in terms of several evaluation metrics(e.g.,ad revenue,ROI,ANLP and KL divergence). |
