Scholarlink

<p>Predicting information diffusion cascades in social networks is crucial, particularly for forecasting cascade sizes. However, existing methods often struggle to handle the complex dynamics of information spread, especially in terms of multi-scale temporal patterns and intricate node interactions. This paper introduces <i>CasMST-KAN</i>, a novel deep learning model designed to address these challenges by integrating advanced techniques for both temporal and spatial modeling. Specifically, we propose a temporal information processing branch that combines <i>multi-scale temporal convolution (MSTC)</i> and long short-term memory (LSTM) to effectively capture dynamic patterns across multiple temporal scales. In parallel, Kolmogorov–Arnold networks (KAN) replace traditional MLPs to more flexibly model complex nonlinear node relationships essential to diffusion. Additionally, a ReLU-based multi-scale linear attention mechanism is introduced to enhance representation learning, improving feature extraction across scales. Extensive experiments on multiple real-world datasets demonstrate that CasMST-KAN significantly outperforms state-of-the-art methods in predicting information diffusion cascades, particularly excelling in long-term cascades where capturing temporal dynamics and deep structural dependencies is crucial. Code and datasets are available at <a href="https://github.com/Slip01/CasMST-KAN">https://github.com/Slip01/CasMST-KAN</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

CasMST-KAN: a cascade prediction model with multi-scale temporal convolution and Kolmogorov–Arnold networks

  • 期刊论文

  • 发表日期：

    2026年3月21日

查看全文
  • Xia Dan,
  • Zhihao Zhang,
  • Yuan Tian,
  • Di Chen

摘要

Predicting information diffusion cascades in social networks is crucial, particularly for forecasting cascade sizes. However, existing methods often struggle to handle the complex dynamics of information spread, especially in terms of multi-scale temporal patterns and intricate node interactions. This paper introduces CasMST-KAN, a novel deep learning model designed to address these challenges by integrating advanced techniques for both temporal and spatial modeling. Specifically, we propose a temporal information processing branch that combines multi-scale temporal convolution (MSTC) and long short-term memory (LSTM) to effectively capture dynamic patterns across multiple temporal scales. In parallel, Kolmogorov–Arnold networks (KAN) replace traditional MLPs to more flexibly model complex nonlinear node relationships essential to diffusion. Additionally, a ReLU-based multi-scale linear attention mechanism is introduced to enhance representation learning, improving feature extraction across scales. Extensive experiments on multiple real-world datasets demonstrate that CasMST-KAN significantly outperforms state-of-the-art methods in predicting information diffusion cascades, particularly excelling in long-term cascades where capturing temporal dynamics and deep structural dependencies is crucial. Code and datasets are available at https://github.com/Slip01/CasMST-KAN.