Scholarlink

<p>Accurate prediction of popularity for pre-cached content entities is a prerequisite for designing popularity-based dynamic edge caching methods that effectively reduce the overhead of resource-limited edge servers and enhance user experience. Existing methods using LSTM and its variants are limited in their ability to capture periodic patterns and peak variations, which often face challenges in accurately predicting popularity. Our work proposes a model, termed LSTNetPro, for popularity prediction. LSTNetPro replaces the convolutional network in LSTNet with a temporal convolutional network and incorporates an attention mechanism to more effectively capture temporal dependencies and peak variations in data, thereby improving popularity prediction performance. Furthermore, system and optimization models to minimize popularity-based energy consumption are constructed. The edge caching problem is formulated as a Markov Decision Process. Guided by LSTNetPro predictions, an Actor-Critic framework, referred to as LSTNetPro-AC, is employed to achieve dynamic caching optimization decisions. Simulations on three real-world datasets demonstrate that LSTNetPro significantly outperforms baseline models such as LSTM, BiLSTM, and LSTNet across four performance metrics, including Root Relative Squared Error and Empirical Correlation Coefficient. Meanwhile, LSTNetPro-AC surpasses six comparative algorithms in popularity-based energy consumption, cache hit rate, time cost, and reward, achieving approximately 90% of the theoretically optimal performance derived from global-optimal dynamic programming.</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

A dynamic edge caching method based on popularity prediction with an improved LSTNet network

  • 期刊论文

  • 发表日期：

    2025年11月22日

查看全文
  • Baoyu Xu,
  • Lingjie Zou,
  • Shuhuai Li,
  • Xiaoyang Kang,
  • Lihua Zhang

摘要

Accurate prediction of popularity for pre-cached content entities is a prerequisite for designing popularity-based dynamic edge caching methods that effectively reduce the overhead of resource-limited edge servers and enhance user experience. Existing methods using LSTM and its variants are limited in their ability to capture periodic patterns and peak variations, which often face challenges in accurately predicting popularity. Our work proposes a model, termed LSTNetPro, for popularity prediction. LSTNetPro replaces the convolutional network in LSTNet with a temporal convolutional network and incorporates an attention mechanism to more effectively capture temporal dependencies and peak variations in data, thereby improving popularity prediction performance. Furthermore, system and optimization models to minimize popularity-based energy consumption are constructed. The edge caching problem is formulated as a Markov Decision Process. Guided by LSTNetPro predictions, an Actor-Critic framework, referred to as LSTNetPro-AC, is employed to achieve dynamic caching optimization decisions. Simulations on three real-world datasets demonstrate that LSTNetPro significantly outperforms baseline models such as LSTM, BiLSTM, and LSTNet across four performance metrics, including Root Relative Squared Error and Empirical Correlation Coefficient. Meanwhile, LSTNetPro-AC surpasses six comparative algorithms in popularity-based energy consumption, cache hit rate, time cost, and reward, achieving approximately 90% of the theoretically optimal performance derived from global-optimal dynamic programming.