Scholarlink

<p>Graph neural networks excel in representation learning on structured graphs, yet many decoupled frameworks ignore the intrinsic link between structure and features and offer limited means to mitigate noisy neighborhood aggregation. To address these limitations, we propose DGNN-SGS, a dual-graph, weight-tied GNN that constructs separate embeddings from the feature and topology spaces for fine-grained decoupling while sharing a single convolution–gating block across views. By integrating feature and topological graphs, we enhance node representation quality. In addition, a spatially-aware gating mechanism is incorporated to adaptively filter neighbor information, yielding not only superior predictive accuracy but also more efficient training compared with prior designs. Extensive experiments on benchmark datasets demonstrate that DGNN-SGS significantly improves node classification performance.</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

Decoupled graph neural network with spatially-aware gating selections

  • 期刊论文

  • 发表日期：

    2025年10月13日

查看全文
  • Xin Meng,
  • ShuXia Lu

摘要

Graph neural networks excel in representation learning on structured graphs, yet many decoupled frameworks ignore the intrinsic link between structure and features and offer limited means to mitigate noisy neighborhood aggregation. To address these limitations, we propose DGNN-SGS, a dual-graph, weight-tied GNN that constructs separate embeddings from the feature and topology spaces for fine-grained decoupling while sharing a single convolution–gating block across views. By integrating feature and topological graphs, we enhance node representation quality. In addition, a spatially-aware gating mechanism is incorporated to adaptively filter neighbor information, yielding not only superior predictive accuracy but also more efficient training compared with prior designs. Extensive experiments on benchmark datasets demonstrate that DGNN-SGS significantly improves node classification performance.