Scholarlink

<p>Online recommendation systems quickly develop personalized recommendations based on users’ historical feedback, thereby improving user experience and increasing platform revenue. Contextual Multi-Armed Bandits (CMAB) model based on reinforcement learning can achieve an effective balance between exploration and utilization, thereby maximizing long-term returns. In this work, we propose a novel CMAB model for online recommendation, which introduces two key innovations: (1) Knowledge Graph-driven Thompson Sampling (KG-TS) that enriches context by constructing a dynamic knowledge graph from user-item interactions to alleviate data sparsity, and (2) Time-Varying Reward Mechanism (TV-RM) that dynamically updates graph edges based on real-time feedback to adapt to non-stationary environments. The integrated algorithm, named KG-TV-TS, is designed to handle sparse and evolving recommendation scenarios. Experiments on three public datasets demonstrate that KG-TV-TS consistently outperforms state-of-the-art bandit algorithms in both recommendation accuracy and cumulative regret, especially under sparse and non-stationary conditions.</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

Online recommendation in non-stationary environments based on knowledge graph enhancement and time-varying reward mechanism

  • 期刊论文

  • 发表日期：

    2026年2月3日

查看全文
  • Jiaxin Li,
  • Jinyun Fang

摘要

Online recommendation systems quickly develop personalized recommendations based on users’ historical feedback, thereby improving user experience and increasing platform revenue. Contextual Multi-Armed Bandits (CMAB) model based on reinforcement learning can achieve an effective balance between exploration and utilization, thereby maximizing long-term returns. In this work, we propose a novel CMAB model for online recommendation, which introduces two key innovations: (1) Knowledge Graph-driven Thompson Sampling (KG-TS) that enriches context by constructing a dynamic knowledge graph from user-item interactions to alleviate data sparsity, and (2) Time-Varying Reward Mechanism (TV-RM) that dynamically updates graph edges based on real-time feedback to adapt to non-stationary environments. The integrated algorithm, named KG-TV-TS, is designed to handle sparse and evolving recommendation scenarios. Experiments on three public datasets demonstrate that KG-TV-TS consistently outperforms state-of-the-art bandit algorithms in both recommendation accuracy and cumulative regret, especially under sparse and non-stationary conditions.