Receding Bid Optimization Method with Real-Time Feedback and Prediction for Sponsored Search Advertising on Taobao

Since the popularity of the Internet, sponsored search advertising has become one of the most vital marketing methods in e-commerce. As one of its cores, the bid optimization problem seeks a bidding strategy that helps advertisers maximize their objectives subject to commercial constraints. We model it as a Markov Decision Process (MDP); however, directly solving such MDPs is inefficient and impractical due to the complexity, computation, and magnitude. Thus, we propose the receding bid optimization method. Then the problem is transformed into a constructed bidding formulation with finite parameters. Meanwhile, we show that parameters can be solved by data-driven linear programming (LP) based on real-time feedback on constraints. Also, the optimality of the bidding formulation is proven. Furthermore, this method must be computation-friendly, considering Taobao's magnitude and reality consumption. Therefore, a down-sampling process is proposed, which achieves over 90% accuracy with only 3% of data. Besides, an auction scale prediction module is added to improve the adaptiveness under dynamic markets. Finally, two-month-long online experiments validate its outstanding performance compared to several state-of-the-art methods. Nowadays, the proposed method has been deployed on Taobao and is used for nearly billions of auctions every day.