HiFI: Hierarchical Fairness-aware Integrated Ranking with Constrained Reinforcement Learning

Yifan Liu, Wei Xia, Weiwen Liu, Menghui Zhu, Weinan Zhang, Ruiming Tang, Yong Yu

Abstract: Integrated ranking is a critical component in industrial recommendation platforms. It combines candidate lists from different upstream channels or sources and ranks them into an integrated list, which will be exposed to users. During this process, to take responsibility for channel providers, the integrated ranking system needs to consider the exposure fairness among channels, which directly affects the opportunities of different channels being displayed to users. Besides, personalization also requires the integrated ranking system to consider the user's diverse preference on different channels besides items. Existing methods are hard to address both problems effectively. In this paper, we propose a <u>Hi</u>erarchical <u>F</u>airness-aware <u>I</u>ntegrated ranking (HiFI) framework. It contains a channel recommender and an item recommender, and the fairness constraint is on channels with constrained RL. We also design a gated attention layer (GAL) to effectively capture users' multi-faceted preferences. We compare HiFI with various baselines on public and industrial datasets, and HiFI achieves the state-of-the-art performance on both utility and fairness metrics. We also conduct an online A/B test to further validate the effectiveness of HiFI.