Scholarlink

<p>Recommender systems (RS) built on deep reinforcement learning (DRL) treat recommendation as a dynamic process, updating their strategies during interaction with the environment based on immediate user rewards. This contrasts with traditional methods, which assume a fixed process and apply a static strategy. However, DRL algorithms can suffer from instability and tend to converge to local optima. Beyond these DRL-specific issues, inefficient rating functions for generating recommendation lists can further degrade overall system performance, particularly the ranking quality of recommender systems. This paper proposes a recommender system called IR<sup>2</sup>SRL, which adopts a DRL actor–critic architecture to improve recommendation ratings. The system employs a time-order method to capture users’ dynamic preferences and implements a weighted Shannon index function. We use a real reward simulation method with a time-order approach and the Facebook algorithm (faiss.IndexHNSWFlat), utilizing the former to simulate real rewards and the latter, which is faster and more accurate, to retrieve similar items. Together, these also reduce training time. We then adopt a combined reward function to enhance both ratings and recommendation diversity, as well as to increase the generalizability of the real reward simulation method under the time-order approach. In addition, we estimate the Target Q-value in a manner that more effectively reduces error variance, enhances model stability, and consequently improves recommendation ratings. Extensive experiments on the MovieLens dataset demonstrate that our proposed model substantially outperforms multiple baselines in ranking quality, efficiency, and diversity. The model yields an absolute gain of + 0.076 in NDCG (an 8.4% relative improvement) and a 42.8% average improvement over baselines, while Precision improves by 33.02% on average. It achieves absolute gains of + 5.523/ + 13.335 in Cumulative NDCG@10/@20 (161%/241% relative to the strongest baseline; 162%/243% on average) and + 0.712/ + 0.889 in Cumulative Precision@10/@20 (10.39%/7.51% relative; 13.59%/9.95% on average), The Proposed model also reduces the average training time per episode by 71.10% and enhances diversity by 0.093 (10.8%), confirming its superior effectiveness, efficiency, and robustness.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

IR2SRL: improving the rating of recommender systems with a deep reinforcement learning approach

  • 期刊论文

  • 发表日期：

    2026年4月28日

查看全文
  • Ali Yarahmadi,
  • Ghasem Azari Arani,
  • Mahdi Esmaeili

摘要

Recommender systems (RS) built on deep reinforcement learning (DRL) treat recommendation as a dynamic process, updating their strategies during interaction with the environment based on immediate user rewards. This contrasts with traditional methods, which assume a fixed process and apply a static strategy. However, DRL algorithms can suffer from instability and tend to converge to local optima. Beyond these DRL-specific issues, inefficient rating functions for generating recommendation lists can further degrade overall system performance, particularly the ranking quality of recommender systems. This paper proposes a recommender system called IR2SRL, which adopts a DRL actor–critic architecture to improve recommendation ratings. The system employs a time-order method to capture users’ dynamic preferences and implements a weighted Shannon index function. We use a real reward simulation method with a time-order approach and the Facebook algorithm (faiss.IndexHNSWFlat), utilizing the former to simulate real rewards and the latter, which is faster and more accurate, to retrieve similar items. Together, these also reduce training time. We then adopt a combined reward function to enhance both ratings and recommendation diversity, as well as to increase the generalizability of the real reward simulation method under the time-order approach. In addition, we estimate the Target Q-value in a manner that more effectively reduces error variance, enhances model stability, and consequently improves recommendation ratings. Extensive experiments on the MovieLens dataset demonstrate that our proposed model substantially outperforms multiple baselines in ranking quality, efficiency, and diversity. The model yields an absolute gain of + 0.076 in NDCG (an 8.4% relative improvement) and a 42.8% average improvement over baselines, while Precision improves by 33.02% on average. It achieves absolute gains of + 5.523/ + 13.335 in Cumulative NDCG@10/@20 (161%/241% relative to the strongest baseline; 162%/243% on average) and + 0.712/ + 0.889 in Cumulative Precision@10/@20 (10.39%/7.51% relative; 13.59%/9.95% on average), The Proposed model also reduces the average training time per episode by 71.10% and enhances diversity by 0.093 (10.8%), confirming its superior effectiveness, efficiency, and robustness.