Response Prediction On Real Time Bidding Via Tensor Factorization

Real time bidding is changing the way ad inventory is bought from media buying to audience buying. Real time bidding improves the advertising effect and becomes the future of digital advertising. Demand-side platforms is the core one in real-time bidding ecosystem, and must have more powerful capacity for quality evaluation, traffic selection and control than ever, so the demand-side platforms also confront more serious challenges in terms of technologies and algorithms than ever. The most different task of the demandside platform is to bid for ad impressions in public auction. Therefore, how to bid is one of their key product strategies, and the bid prices directly determine their profits. Furthermore, the prediction quality of click-through rate and conversion rate directly impacts on the quality of the bid price. Therefore, how to accurately estimate click-through rate and conversion rate becomes one of the core issues to be addressed for demand-side platform in real-time bidding. This thesis focuses on the methods and technologies of demand-side platforms which will improve the prediction performance of the click-through rate and conversion rate in real-time bidding systems.When addressing the response prediction, the demand-side platform faces with many daunting challenges mainly coming from the following four aspects, high data sparsity of user feedback, high speed requirement of real time task, more complexity of the interactions between the user, context and ad, serious class imbalance in the historical feedback.Therefore, this thesis focuses on the methods and strategies to address these four challenges.In particular, our research work mainly involves the following four aspects:Firstly, in order to alleviate the serious data sparsity problem, this thesis researches on the solution of the integration of heterogeneous information based on tensor factorization models. At first, this thesis attempts to extend user features from user-generated tags.And then based on tensor factorization models, this thesis presents different integration strategies for different heterogeneous information according to their characteristics. Finally, we propose an integrative framework to integrate different heterogeneous information into tensor decomposition models together and improve the accuracy and reliability of response prediction. We apply the strategies into matrix and tensor factorization model and conduct comparison experiments on three specified data sets. The experimental results show that the heterogeneous information integration technology significantly improves the prediction performance.Secondly, traditional Tucker decomposition model can’t satisfy the requirements of real time bidding both on prediction performance and runtime. Aiming to reduce its computational runtime, this thesis researches on response prediction method based on latent cube factorization model. We theoretically analyze the relationship between singular value decomposition and matrix factorization model, and propose the latent cube factorization model. This model is directly trained on the observation data through using a deformation of truncated higher order singular value decomposition. Therefore, this model reduces runtime complexity from cubic to linear runtime based on the number of latent factors and simultaneously guarantees better prediction performance. The latent cube factorization model has few parameters and is easy to be implemented. The experimental results show that the latent cube factorization model remains the better prediction quality than CP decomposition and faster runtime than Tucker decomposition.Thirdly, in order to obtain fully pairwise interaction relation between the three objects, this thesis researches on fully coupled interaction tensor factorization model. We analyze the basic idea of pairwise interaction tensor factorization model and its limitation in solving response prediction. Then, we present fully coupled interaction tensor decomposition model which finally overcomes these limitations. This model learns two different latent factors for each object corresponding to each interaction with one of the other two ones. Therefore, this model improves the prediction performance and simultaneously ensures the runtime complexity is not higher than linear time of the number of latent factors.Finally, in order to relieve the impact of class imbalance on prediction quality, this thesis researches on the method of response prediction based on triple-wise learning via utilizing the click and converse response data simultaneously. We formulate response prediction as a ranking problem and take the correct ranking of different impression classes as the new optimization objective. Then, we propose a novel ranking optimization strategy named triplet-wise learning based on optimizing the proper order of random triples composed of conversion, click-only and non-click. This learning strategy supports conversion prediction via using click feedback information and reduces the effect of clicking noise via conversion feedback information. Because our method no longer takes minimizing the estimation error of sample instant as the optimization objective, it effectively relieves the impact of the imbalance between positive and negative samples on model training.In short, in order to address click through rate and conversion rate prediction problems using tensor decomposition models for the demand-side platform in real time bidding, This thesis mainly focuses on tackling various tough challenges and proposes appropriate methods and technological solutions corresponding to different difficulties to effectively tackle them and achieves better performance both in prediction quality and in runtime complexity.