Scholarlink

<p>Existing long-term sequential recommendations ignore the capture of real-time user preferences, which leads to poor recommendation accuracy. To solve this problem, this work proposes a decoupled long-term sequential recommendation model leveraging competitive mechanism (DSR-CM). By introducing a multi-head flow-attention mechanism, the user preference information flow competes spontaneously under the constraints of the competition mechanism, which preserves the learning ability of traditional dot-product attention at the level of linear computational complexity, and captures the dynamic preference relations in the sequence of user behaviours more efficiently. Meanwhile, using decoupled computational position encoding more accurately captures the sequential relationship between user behaviours and models the user preference trends. Extensive experimental studies are carried out on three real-world datasets, where DSR-CM outperforms the existing state-of-the-art methods in terms of both effectiveness and efficiency, it accurately captures user’s dynamically changing preferences and latest trends, and explores new ways to enhance the adaptability and accuracy of the recommender system. The implementation code is available online at <a href="https://github.com/cyxg7/DSR-CM">https://github.com/cyxg7/DSR-CM</a>.</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

DSR-CM: decoupled long-term sequential recommendation model leveraging competitive mechanism

  • 期刊论文

  • 发表日期：

    2025年2月18日

查看全文
  • Shaoguo Cui,
  • Xingyu Li

摘要

Existing long-term sequential recommendations ignore the capture of real-time user preferences, which leads to poor recommendation accuracy. To solve this problem, this work proposes a decoupled long-term sequential recommendation model leveraging competitive mechanism (DSR-CM). By introducing a multi-head flow-attention mechanism, the user preference information flow competes spontaneously under the constraints of the competition mechanism, which preserves the learning ability of traditional dot-product attention at the level of linear computational complexity, and captures the dynamic preference relations in the sequence of user behaviours more efficiently. Meanwhile, using decoupled computational position encoding more accurately captures the sequential relationship between user behaviours and models the user preference trends. Extensive experimental studies are carried out on three real-world datasets, where DSR-CM outperforms the existing state-of-the-art methods in terms of both effectiveness and efficiency, it accurately captures user’s dynamically changing preferences and latest trends, and explores new ways to enhance the adaptability and accuracy of the recommender system. The implementation code is available online at https://github.com/cyxg7/DSR-CM.