Distributed dominance graph-based neural multi-objective evolutionary strategy for sponsored search real-time bidding

Abstract

In industrial online advertising, real-time bidding optimization has brought substantial profits. Existing methods usually set fixed bids for advertisers which causes collusion among advertisers, resulting in unsatisfactory performance. In this work, by considering both advertisers’ and the platform’s objectives, we formulate the problem as a multiobjective optimization problem and propose a neural evolutionary strategy-based framework to learn the optimal bidding strategy. Since the objectives of advertisers and the platform are conflicting to some extent, it is crucial to properly balance them. A straightforward idea is to simply add up the objectives. However, this could limit the diversity of solutions, leading to suboptimal performance. To this end, we innovatively develop a distributed dominance graph-based neural Pareto multiobjective evolutionary strategy, which can learn a better spread of solutions to improve the model performance. In addition, it can distribute the total computational cost among thousands of workers, improving computational efficiency. We conduct a suite of offline experiments as well as the standard online A/B test on the Taobao platform. The experimental results indicate that our method can significantly improve the profits of both advertisers and the platform.