Modeling And Decision Optimization In Real-time Bidding Advertising

Nowadays,online advertising,a.k.a.digital marketing,whose goal is to bridge the gap between merchandise sellers and common customers through the online media,has become the main source for the Internet-based business platform,such as ecommerce websites,search engines and social media.To achieve the goal of digital marketing,computational advertising provides a programmatic approach with scalable,computational and even learnable methodologies via rapidly increasing computing capability.Emerged from 2009,real-time bidding(RTB)is the most widely adopted online advertising paradigm which essentially facilitates buying an individual ad impression in real time while it is still being generated from a user's visit.On one side,RTB scales up the online buying process by aggregating abundant media inventories and supporting flexible marketing strategies.On the other side,it also enables dynamic user targeting via computational approaches.In this cutting-edge frontier area,many researchers have devoted lots of efforts to solve the emerging challenges from these three fundamental aspects of RTB,along with the other considerations of online advertising.From the view of techniques,the solutions cover a wide range of research fields such as information retrieval,machine learning,optimization,economics and game theory.In this thesis,we take a comprehensive modeling view to tackle the main problems of RTB and deal with several other challenges in the RTB ecosystem.We define the tasks of modeling in RTB advertising,and analyze the problems hidden in the existing methods,and finally propose our solutions from five perspectives as bellow.Cost Estimation For bid landscape forecasting,we proposed an auto-regressive deep model which is the first work considering sequential feature dependency with survival analysis for modeling market price distributions.Thus it does not need to make any assumptions about base distribution forms of bid landscape and can generally predict flexible market price distributions.Utility Prediction The existing methods for user response prediction usually regard the problem as a binary classification task,which may not be optimal in RTB scenario.We propose a new objective function for directly optimizing advertiser profits when conducting ad campaigns.The derived model has significantly improved the advertising utility since our method is cost-sensitive which can largely save the budget and allocate for higheffective ad impressions with low costs.Bid Optimization The researchers often optimize the bid function and user response prediction function w.r.t.the isolated objectives.They have never combined these two problems.However,it is more effective considering the overall optimization procedure in RTB scenario.Thus it is eager to optimize through a comprehensive approach to improve the model efficiency.We propose a unified objective function of advertiser's profits and optimize it from three aspects,i.e.,user response prediction,bid landscape forecasting and bid optimization,in the whole framework,which has significantly improved the advertising performance.User Modeling In RTB or other online information systems,user modeling is a key component for better user interest capturing and behavior pattern mining.Thus it provides some valuable clues for subsequent decision making in information systems,such as better advertising and recommendation,etc.To model the abundant yet multi-facet user behaviors in the online system,we propose a lifelong user modeling framework with hierarchical and periodic memory network,to store the user behavior patterns along her long-term history of online activities.It also conducts a comprehensive solution for user behavior prediction.Conversion Attribution To better understand user behaviors and direct the subsequent ad delivery,we also conduct a deep neural network for multi-touch conversion attribution.The proposed deep model improves the conversion prediction in a large margin and it also provides comprehensive conversion attributions for each user behavior sequence.To quantitatively benchmark attribution models,we also propose a novel yet practical attribution evaluation scheme through the proxy of budget allocation(under the estimated attributions)over ad channels.Besides,both the academic community and industrial applications have demonstrated the effectiveness and efficiency of the modeling perspective and the corresponding solutions in this thesis.We not only evaluate our models against state-of-the-arts on several large-scale datasets,but also conduct online experiments on real-world advertising platforms.Moreover,the proposed methodologies can also facilitate many more widely spread applications in other research areas such as recommender systems and search engines.