Scholarlink

<p>Dynamic Graph Neural Networks (DGNNs) have emerged as a powerful framework for modeling the structural and temporal evolution of dynamic systems, with extensive applications in fields such as finance and communication networks. However, existing methods still face challenges due to the limited expressive capability of time encoding functions with fixed forms and the lack of temporal awareness in local neighborhood aggregation, which restricts the fine-grained modeling of dynamic dependencies. In this paper, we propose a <b>L</b>earnable <b>T</b>emporal <b>F</b>unction-based <b>Dy</b>namic <b>G</b>raph Neural Network with Dual-channel Encoding (LTFDyG). This method adopts a hierarchical encoding architecture to systematically fuse temporal and structural information from local to global levels. First, a spectral-entropy-guided learnable temporal function blends Fourier and spline bases to adaptively encode time intervals into a continuous embedding space, forming an expressive temporal feature channel. Second, the learnable temporal function incorporates a temporal modulation mechanism into the neighbor co-occurrence frequency statistics, creating a temporal-aware neighbor interaction channel that captures temporal differences in local neighborhoods. Then, features from the two channels are fused with node and edge features into a unified embedding space. Finally, a Transformer architecture, utilizing both self-attention and cross-attention mechanisms, models the global structural and temporal dependencies. Experimental results on six real-world datasets and seven baseline methods demonstrate the superior performance of LTFDyG across both the link prediction and node classification tasks. The source code has been made available at <a href="https://github.com/Pacino-118/LTFDyG/tree/master">https://github.com/Pacino-118/LTFDyG/tree/master</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

LTFDyG: A learnable temporal function-based dynamic graph neural network with dual-channel encoding

  • 期刊论文

  • 发表日期：

    2026年5月25日

查看全文
  • Xinzhi Shi,
  • Chao Li,
  • Xingshuo Han,
  • Shihe Su,
  • Junyan Wu

摘要

Dynamic Graph Neural Networks (DGNNs) have emerged as a powerful framework for modeling the structural and temporal evolution of dynamic systems, with extensive applications in fields such as finance and communication networks. However, existing methods still face challenges due to the limited expressive capability of time encoding functions with fixed forms and the lack of temporal awareness in local neighborhood aggregation, which restricts the fine-grained modeling of dynamic dependencies. In this paper, we propose a Learnable Temporal Function-based Dynamic Graph Neural Network with Dual-channel Encoding (LTFDyG). This method adopts a hierarchical encoding architecture to systematically fuse temporal and structural information from local to global levels. First, a spectral-entropy-guided learnable temporal function blends Fourier and spline bases to adaptively encode time intervals into a continuous embedding space, forming an expressive temporal feature channel. Second, the learnable temporal function incorporates a temporal modulation mechanism into the neighbor co-occurrence frequency statistics, creating a temporal-aware neighbor interaction channel that captures temporal differences in local neighborhoods. Then, features from the two channels are fused with node and edge features into a unified embedding space. Finally, a Transformer architecture, utilizing both self-attention and cross-attention mechanisms, models the global structural and temporal dependencies. Experimental results on six real-world datasets and seven baseline methods demonstrate the superior performance of LTFDyG across both the link prediction and node classification tasks. The source code has been made available at https://github.com/Pacino-118/LTFDyG/tree/master.