Reward Constrained Interactive Recommendation with Natural Language Feedback
This work addresses the challenge of balancing exploration and user preference adherence in text-based interactive recommendation, which is incremental as it builds on existing RL methods by adding constraints.
The paper tackles the problem of interactive recommendation systems violating user preferences when exploring new items, by proposing a constraint-augmented reinforcement learning framework that incorporates historical natural language feedback to detect and avoid such violations, resulting in consistent improvements over standard RL methods.
Text-based interactive recommendation provides richer user feedback and has demonstrated advantages over traditional interactive recommender systems. However, recommendations can easily violate preferences of users from their past natural-language feedback, since the recommender needs to explore new items for further improvement. To alleviate this issue, we propose a novel constraint-augmented reinforcement learning (RL) framework to efficiently incorporate user preferences over time. Specifically, we leverage a discriminator to detect recommendations violating user historical preference, which is incorporated into the standard RL objective of maximizing expected cumulative future rewards. Our proposed framework is general and is further extended to the task of constrained text generation. Empirical results show that the proposed method yields consistent improvement relative to standard RL methods.