Research On Predictive Modeling And Bidding Optimization For Real-time Bidding Advertising

With the rapid evolution of internet technology,online advertising has become the mainstream of digital marketing.Serving as a link between businesses and users,online advertising not only provides users with high-quality information,but also generates substantial economic benefits for enterprises.To meet the increasingly refined marketing demands and balance the supply-demand relationship in the advertising market,the emerging real-time bidding(RTB)trading mechanism has brought disruptive changes to the delivery technology and transaction mode for online advertising.Different from the traditional guaranteed delivery paradigm,RTB integrates vast advertising resources and empowers advertisers to purchase any single ad display opportunity through programmatic bidding in real-time.Since its emergence in 2009,RTB has gradually captured the market with its indispensable advantages,including personalized targeting and optimized resource allocation.In order to guarantee the cost-effectiveness of digital marketing,advertisers must not only accurately estimate both the advertising utility and cost,but also develop efficient bidding strategies to calculate appropriate bids to participate in realtime auctions.In this thesis,we concentrate on predictive modeling and bidding optimization challenges in the RTB advertising scenario,and we propose innovative research solutions as follows.1)Most existing methods in RTB advertising treat utility and cost estimation as independent tasks to be optimized separately.However,through quantitative data analysis,we reveal a strong task-relatedness between them and demonstrate that optimizing each of them independently can not achieve the goal of global optimization.To address the challenge,we propose a multitask learning framework that jointly optimizes the two critical estimation tasks in an end-toend manner.The framework empowers shared representations through task-related knowledge transfer in order to improve generalization and prediction performance.For cost estimation,most current research relies on statistical analysis to model the market price by making prior distribution assumptions.However,such heuristic empirical assumptions are too restrictive to adapt dynamically to the complex RTB market.To solve this problem,we propose to cast the cost estimation task as a multi-class classification task to predict the probabilities of all discrete market prices.In this way,the model can utilize the features of the bid request granularity for personalized estimations,without the need for prior distribution assumptions.Moreover,due to the fact that advertisers can only observe the supervised signals of clicks and market prices when they win the bid,there is a severe sample selection bias problem in RTB.To alleviate this bias issue,we utilize the abundant bid price information in the full bidding space and introduce an auxiliary task of predicting the winning probability to perform unbiased learning.2)To achieve the goal of maximizing advertisers’ profits in RTB advertising,we propose an effective bidding strategy called Adaptive ROI-aware Bidding.Specifically,we conduct a comprehensive analysis of the influence of budget and available auction volume on the marginal benefits and propose a quantitative criterion to determine whether the budget is sufficient.This allows the bidding strategy to be dynamically adapted to various budget settings.In the cases of insufficient budget,we theoretically prove that the advertiser’s profit can only be maximized when the overall expected cost of participating in all auctions exactly equals the budget.Based on this optimization condition,advertisers can efficiently solve the optimal bid coefficient in the bidding function.In addition,we propose a novel bidding function based on return on investment(ROI).Unlike most existing bidding functions that bid based on estimated utility,ROI-aware bidding additionally takes the estimated cost information into consideration,thereby improving the profit in a more cost-effective manner.3)Building on the two aforementioned research works,we propose to enhance the structural design and optimization objectives for multi-task learning in RTB.In the aspect of improving the structural design of shared networks,we propose a novel network called Progressive Dropout Customized Mixture-of-Experts.Specifically,the network explicitly separates task-specific experts from task-shared experts to reduce information interference caused by task conflicts and alleviate negative transfer problems.Furthermore,we apply progressive dropout regularization to all expert units to avoid problems of fusion weight imbalance and expert collapse.From the view of improving optimization objectives in multi-task learning,we propose a Pareto optimization method.Specifically,we cast the original multi-task learning problem as a multi-objective optimization problem in order to find a Pareto optimal solution that has good trade-offs among multiple tasks.In addition,we propose to take the key performance indicators(KPI)of advertisers as the overall optimization objective,and unify the optimization of two crucial prediction models and the bidding strategy to achieve global optimization.To validate the effectiveness of the aforementioned research work,we conducted extensive experiments on several large-scale public datasets in the RTB scenario.The experimental results show that the proposed method achieved significant improvements in several key evaluation metrics,such as prediction accuracy and bidding effectiveness.The research findings have been recognized by both academia and industry,and are promising to promote relevant theoretical and technological research in computational advertising.