Optimizing Multiple Performance Metrics with Deep GSP Auctions for E-commerce Advertising

In online e-commerce advertising, the ad platform relies on auction mechanisms to optimize multiple performance metrics, such as user experience, advertiser utility and platform revenue. However, the state-of-the-art ad auctions usually focus on single objective, e.g., either social welfare or revenue, which are not suitable for e-commerce advertising with various, dynamic, conflicting and hardly estimated ad performance metrics. In this paper, we propose a Deep GSP auction mechanism for display advertising, which leverages deep learning techniques to design a new rank score under the celebrated GSP auction framework. Such a deep rank score can well capture the potential contribution of each ad to the optimization objective. We implement this rank score via a deep neural network model under the constraints of monotone allocation and smooth transition. The requirement of monotone allocation enables the deep GSP auction to achieve nice game theoretical equilibrium, and the smooth transition constraint guarantees the advertiser utility would not fluctuate too much when the auction changes for a new optimization objective. We conducted comprehensive experiments in both offline simulation and online A/B test. The evaluation results demonstrate the effectiveness of the Deep GSP auction in terms of various ad performance metrics, compared to the state-of-the-art ad auction mechanisms.