Abstract: To address the issues of structural unfairness and classification accuracy in the Graph Attention Network (GAT) for node classification tasks, a novel Graph Neural Network node classification model was proposed, incorporating clustering coefficients. First, the clustering coefficients of neighboring nodes were introduced as structural information and combined with trainable weight parameters, which enhanced the representation of the topological structure in the attention mechanism. Then, feature scaling techniques were employed to optimize node embeddings, and residual connections were incorporated to mitigate the risk of feature over-smoothing. Experimental results on six real-world datasets demonstrated that the proposed model outperformed mainstream models, such as GIN (Graph Isomorphism Network) and GOAT (Graph Ordering Attention Networks), in classification accuracy. For instance, compared to the baseline model GAT, the classification accuracy of the proposed model improved by 4 percentage points on the Cora dataset, the structural bias was reduced from 0.31% to 0.11%, and the classification accuracy of isolated nodes improved by 3.69 percentage points. In conclusion, the proposed model not only achieved significant improvements in classification performance but also demonstrated superiority in structural fairness and stability.

Key words: node classification, attention mechanism, clustering coefficient, topological information, structural fairness

摘要： 针对图注意力模型（GAT）在节点分类任务中存在的结构不公平性和分类准确性问题，提出了一种引入聚类系数的图神经网络节点分类模型。首先，该模型通过引入邻居节点的聚类系数作为结构信息，结合可训练的权重参数，增强了拓扑结构在注意力机制中的表达能力；其次，采用特征缩放优化节点嵌入，并加入残差连接以减轻特征过平滑的风险。在6个真实数据上的实验结果表明，该模型的分类准确率超过GIN(Graph Isomorphism Network)、GOAT(Graph Ordering Attention Networks)等主流模型。例如，相较于基线模型GAT，所提模型的分类准确率在Cora数据集上提升了4个百分点，结构性偏差从0.31%降低到0.11%，孤立节点分类的准确率提升了3.69个百分点。综上，所提模型不仅在分类性能上取得显著提升，更在结构公平性与稳定性方面展现出优越性。

关键词: 节点分类, 注意力机制, 聚类系数, 拓扑信息, 结构公平性