Scholarlink

<p>Popularity bias remains a critical challenge in recommendation systems, where a small subset of popular items dominates user exposure, reducing content diversity and user satisfaction. Existing debiasing approaches–such as post-processing, causal inference, and adversarial learning–have shown effectiveness, yet most rely on static models and overlook the dynamic nature of user preferences and feedback loops. Moreover, current dynamic methods focus narrowly on novelty enhancement, failing to address the combined impact of data sparsity and model bias. To overcome these limitations, we propose DRPRCDA (<Emphasis Type="Underline">D</Emphasis>ynamic <Emphasis Type="Underline">R</Emphasis>elief <Emphasis Type="Underline">P</Emphasis>opularity <Emphasis Type="Underline">R</Emphasis>ecommendation <Emphasis Type="Underline">C</Emphasis>onsidering <Emphasis Type="Underline">D</Emphasis>iversity and <Emphasis Type="Underline">A</Emphasis>ccuracy), an end-to-end framework that integrates graph-based representation learning and bias-aware reinforcement learning. Specifically, a Light Graph Convolution Network module mitigates data sparsity by generating high-quality embeddings, which are then used by a reinforcement learning agent equipped with a dual-objective reward function to balance recommendation accuracy and diversity. A dynamic regularization mechanism further reduces the influence of popular items by adaptively adjusting their contribution to Q-value updates. Extensive experiments demonstrate that DRPRCDA effectively alleviates popularity bias while maintaining or improving recommendation performance in dynamic environments. This work provides a comprehensive solution for fair, diverse, and adaptive recommendation in real-world scenarios.</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

Dynamic debiasing for popularity bias in recommendation systems: a reinforcement learning approach

  • 期刊论文

  • 发表日期：

    2026年2月24日

查看全文
  • Feng Yang,
  • Xiuli Geng

摘要

Popularity bias remains a critical challenge in recommendation systems, where a small subset of popular items dominates user exposure, reducing content diversity and user satisfaction. Existing debiasing approaches–such as post-processing, causal inference, and adversarial learning–have shown effectiveness, yet most rely on static models and overlook the dynamic nature of user preferences and feedback loops. Moreover, current dynamic methods focus narrowly on novelty enhancement, failing to address the combined impact of data sparsity and model bias. To overcome these limitations, we propose DRPRCDA (Dynamic Relief Popularity Recommendation Considering Diversity and Accuracy), an end-to-end framework that integrates graph-based representation learning and bias-aware reinforcement learning. Specifically, a Light Graph Convolution Network module mitigates data sparsity by generating high-quality embeddings, which are then used by a reinforcement learning agent equipped with a dual-objective reward function to balance recommendation accuracy and diversity. A dynamic regularization mechanism further reduces the influence of popular items by adaptively adjusting their contribution to Q-value updates. Extensive experiments demonstrate that DRPRCDA effectively alleviates popularity bias while maintaining or improving recommendation performance in dynamic environments. This work provides a comprehensive solution for fair, diverse, and adaptive recommendation in real-world scenarios.