ACPR: Proactive Recommendation under User Acceptance Constraints

Recommender systems typically prioritize immediate satisfaction, inadvertently reinforcing the Filter Bubble and limiting exploration. Proactive Recommendation addresses this by enhancing interest in a target item through a sequence of intermediate items. Prior works typically assume that users will passively accept all intermediate items, and primarily focus on the interest uplift toward the target item. However, in practice, if an intermediate item deviates significantly from a user’s interest, the user is likely to actively abandon the current session, resulting in the complete failure of the target item’s exposure. To address this, we propose the Acceptance-Constrained Proactive Recommendation (ACPR) framework, formulating proactive recommendation as a Constrained Markov Decision Process (CMDP) with user acceptance as a safety constraint. Instead of optimizing cumulative rewards related to the target, ACPR employs a step-wise greedy strategy to maintain every intermediate item within the user’s current acceptance region, securing a feasible sequence for effective target exposure. Specifically, we instantiate the policy using a Large Language Model (LLM) to leverage its rich semantic knowledge to infer natural and logical connections between diverse items. To solve the constrained optimization problem, we integrate Reward Constrained Policy Optimization (RCPO) with Group Relative Policy Optimization (GRPO) to balance the proactive objective and acceptance constraint. Extensive experiments demonstrate that ACPR effectively improves users’ interest in target items while maintaining high intermediate acceptance compared to state-of-the-art methods. The advantage of ACPR is particularly evident in scenarios with larger interest gaps, especially in sparse data scenarios where limited user history makes preference understanding more difficult. Code is available at: https://github.com/YexuanChe00/ACPR