Scholarlink

<p>Recommendation systems play a critical role in enhancing user experiences across various online platforms by delivering personalized content. However, traditional learning algorithms, including Restricted Boltzmann Machines (RBMs), often face significant challenges when dealing with large data volumes and sparse data distributions, which can lead to instability and poor performance. To address these shortcomings, this paper presents two novel algorithms: RBM-Q-Learning 1 and RBM-Q-Learning 2. These algorithms introduce advanced mechanisms for state representation and action selection, specifically designed to improve stability and robustness during the training process. The effectiveness of the proposed methods is evaluated through extensive experiments on three datasets from the MovieLens platform—MovieLens 100K, 1M and 10M. Performance is measured using MAE, RMSE, HR, ARHR, Diversity, and Novelty metrics. Our findings demonstrate that the proposed algorithms achieve significantly improved stability and performance, particularly in handling large-scale and sparse data, thus offering a more reliable solution compared to conventional approaches.</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

Q-learning improved RBM for rate prediction in recommendation systems

  • 期刊论文

  • 发表日期：

    2025年6月1日

查看全文
  • Arif Yelği

摘要

Recommendation systems play a critical role in enhancing user experiences across various online platforms by delivering personalized content. However, traditional learning algorithms, including Restricted Boltzmann Machines (RBMs), often face significant challenges when dealing with large data volumes and sparse data distributions, which can lead to instability and poor performance. To address these shortcomings, this paper presents two novel algorithms: RBM-Q-Learning 1 and RBM-Q-Learning 2. These algorithms introduce advanced mechanisms for state representation and action selection, specifically designed to improve stability and robustness during the training process. The effectiveness of the proposed methods is evaluated through extensive experiments on three datasets from the MovieLens platform—MovieLens 100K, 1M and 10M. Performance is measured using MAE, RMSE, HR, ARHR, Diversity, and Novelty metrics. Our findings demonstrate that the proposed algorithms achieve significantly improved stability and performance, particularly in handling large-scale and sparse data, thus offering a more reliable solution compared to conventional approaches.