Abstract: The Under-Reachability Problem (Under-Reach) was recognized as a critical limitation in graph neural network (GNN) when processing graphs with sparsely and unevenly distributed labeled nodes, where distant unmarked nodes were prevented from receiving effective supervision signals due to topological constraints, consequently degrading model generalization. To address limitations in existing methods including over-smoothing, high computational complexity, and noise sensitivity, a graph neural network framework for topology semantic dual-domain collaboration (TriMix) was proposed through three key innovations. 1) A dynamic mixing ratio mechanism was designed to adaptively adjust weights between pseudo-labels and ground-truth labels across training epochs, prioritizing ground-truth for early-stage stability while gradually incorporating high-confidence pseudo-labels for decision boundary expansion. 2) A topology-semantic dual-domain collaborative node-weighted sampling strategy was constructed by integrating node degree, PageRank values, and feature similarity to quantify node importance, optimizing information propagation paths to enhance reachability of low-centrality nodes. 3) A contrastive learning module was implemented with a triple-tiered negative sampling strategy (category-driven, feature-similarity weighted, and pseudo-label guided) to refine discriminability in the embedding space. Experimental results on benchmark datasets (cora, pubmed, etc.) demonstrated that an accuracy improvement of 2.1%–4.4% in node classification was achieved by TriMix over baseline models such as Graph Convolutional Network (GCN) and Graph Attention Network (GAT), with enhancements also observed in both F1-score and generalization capability. The framework was proven to significantly improve learning efficiency for sparsely labeled graph data through dual-domain optimization of topological and semantic features, providing a robust solution for node classification in complex graph structures.

Key words: Graph Neural Network(GNN), Under-Reachability Problem (Under-Reach), dynamic hybrid, dual-domain collaborative node weighted sampling, contrastive learning

摘要： 图神经网络(GNN)在处理标记节点稀疏且分布不均的图数据时，普遍面临欠可达问题(Under-Reach)，即远端未标记节点因拓扑限制难以接收有效监督信号，导致模型泛化能力受限。现有方法虽能部分缓解这一问题，但仍存在过度平滑、计算复杂度高、噪声敏感等局限性。因此，提出一种拓扑语义双域协同的图神经网络框架(TriMix)，通过三方面改进解决上述挑战。1)动态混合比例机制。基于训练轮次自适应调整伪标签与真实标签的混合权重，在训练初期依赖真实标签稳定收敛，后期逐步引入高置信伪标签以扩展决策边界；2)拓扑-语义双域协同的节点加权采样策略。融合节点度、PageRank值和特征相似性，量化节点重要性并优化信息传播路径，提升低中心性节点的可达性；3)对比学习模块。通过类别驱动、特征相似性加权与伪标签引导的三级负样本生成策略，优化嵌入空间的正负样本区分度，增强模型对未标记数据的语义理解。在cora、pubmed等经典数据集上的实验结果表明，TriMix的节点分类准确率较图卷积网络(GCN)、图注意力网络(GAT)等基线模型提升2.1%~4.4%，F1评分与泛化能力均有所提升。该框架通过拓扑结构与语义特征的双域协同优化，显著提升了稀疏标记图数据的学习效率，为复杂图结构中的节点分类任务提供了思路。

关键词: 图神经网络, 欠可达, 动态混合, 双域协同的节点加权采样, 对比学习