Maximizing Cumulative User Engagement in Sequential Recommendation: An Online Optimization Perspective

• DOI: 10.48550/arxiv.2006.04520
• Publication Date: 2020-01-01
• Authors: Yifei Zhao; Yuhang Zhou; Mingdong Ou; Huan Xu; Nan Li
• Keywords: User Engagement

To maximize cumulative user engagement (e.g. clicks) in sequential recommendation, it is often needed to tradeoff two potentially conflicting objectives, that is, pursuing higher immediate (e.g., click-through rate) and encouraging browsing (i.e., more items exposured). Existing works study these tasks separately, thus tend result sub-optimal results. In this paper, we problem from an online optimization perspective, propose a flexible practical framework explicitly longer length high engagement. Specifically, by considering as actions, user's requests states leaving absorbing state, formulate each behavior personalized Markov decision process (MDP), the of maximizing reduced stochastic shortest path (SSP) problem. Meanwhile, with quit probability estimation, shown SSP can be efficiently solved via dynamic programming. Experiments on real-world datasets demonstrate effectiveness proposed approach. Moreover, approach deployed at large E-commerce platform, achieved over 7% improvement clicks.