Scholarlink

<p>With the rapid development of multimedia systems, personalized recommendation plays a crucial role in enhancing user experience across applications like multimedia streaming and social media computing. Multi-Layer Perceptron (MLP)-based sequential recommendation (SR) aims to model sequential dependencies of user-item interactions using stacked MLP layers. However, it suffers from two limitations. Firstly, it is prone to catastrophic forgetting, which refers to losing previously acquired knowledge when learning new information. Secondly, there is often an imbalance between global and local user interests, along with a limitation in capturing more nuanced semantics. To this end, we pioneer a novel Kolmogorov-Arnold Networks-based paradigm for sequential Recommendation, called KANRec, which aims to better model global and local interests, making it a powerful alternative to traditional MLP-based methods. Specifically, KANRec consists of a global interest module and a local interest module. The two modules’ input representation is generated by integrating the item ID embedding and feature embedding through a KAN layer. In the global interest module, KANRec captures global dependencies from both sequence and channel dimensions and mitigates the catastrophic forgetting problem by using local updates of spline functions. In the local interest module, we propose an innovative Gated Residual Convolutional Kolmogorov-Arnold Networks (GRCKAN) to control the contribution of local information while capturing hidden local features in user-item interactions via edge-level learning. Extensive experiments show the effectiveness of KANRec. The source code and data preprocessing scripts are publicly available at <a href="https://github.com/lllllaaaaa233/KANRec">https://github.com/lllllaaaaa233/KANRec</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

Enhancing global and local interests fusion based on Kolmogorov-Arnold networks for sequential recommendation

  • 期刊论文

  • 发表日期：

    2026年2月3日

查看全文
  • Yili Xu,
  • Xiaohan Fang,
  • Jibing Gong,
  • Yi Zhao,
  • Shijie Wei,
  • Peng Fu,
  • Liping Lv

摘要

With the rapid development of multimedia systems, personalized recommendation plays a crucial role in enhancing user experience across applications like multimedia streaming and social media computing. Multi-Layer Perceptron (MLP)-based sequential recommendation (SR) aims to model sequential dependencies of user-item interactions using stacked MLP layers. However, it suffers from two limitations. Firstly, it is prone to catastrophic forgetting, which refers to losing previously acquired knowledge when learning new information. Secondly, there is often an imbalance between global and local user interests, along with a limitation in capturing more nuanced semantics. To this end, we pioneer a novel Kolmogorov-Arnold Networks-based paradigm for sequential Recommendation, called KANRec, which aims to better model global and local interests, making it a powerful alternative to traditional MLP-based methods. Specifically, KANRec consists of a global interest module and a local interest module. The two modules’ input representation is generated by integrating the item ID embedding and feature embedding through a KAN layer. In the global interest module, KANRec captures global dependencies from both sequence and channel dimensions and mitigates the catastrophic forgetting problem by using local updates of spline functions. In the local interest module, we propose an innovative Gated Residual Convolutional Kolmogorov-Arnold Networks (GRCKAN) to control the contribution of local information while capturing hidden local features in user-item interactions via edge-level learning. Extensive experiments show the effectiveness of KANRec. The source code and data preprocessing scripts are publicly available at https://github.com/lllllaaaaa233/KANRec.