Meta-CRS: A Dynamic Meta-Learning Approach for Effective Conversational Recommender System

• DOI: 10.1145/3604804
• Publication Date: 2023-06-17T09:15:43Z
• Authors: Yuxin Ni; Yunwen Xia; Hui Fang; Chong Long; Xinyu Kong; Daqian Li; Dong Yang; Jie Zhang
• Keywords: Ask price

Conversational recommender system (CRS) enhances the by acquiring latest user preference through dialogues, where an agent needs to decide “whether ask or recommend”, “which attributes ask”, and items recommend” in each round. To explore these questions, reinforcement learning is adopted most CRS frameworks. However, existing studies somewhat ignore consider connection between previous rounds current round of conversation, which might lead lack prior knowledge inaccurate decisions. In this view, we propose facilitate connections different conversations a dialogue session deep transformer-based multi-channel meta-reinforcement learning, so that can action/decision based on states, actions, their rewards. Besides, better utilize user’s historical preferences, more dynamic personalized graph structure support conversation module recommendation module. Experiment results five real-world datasets online evaluation with real users industrial environment validate improvement our method over state-of-the-art approaches effectiveness designs.