SLRNode: node similarity-based leading relationship representation layer in graph neural networks for node classification

Fuchuan Xiang, Yao Xiao, Fenglin Cen, Ji Xu

Abstract: For semi-supervised node classification in graph neural network models (GNNs), the representativeness of pseudo-labeled nodes greatly influences the ultimate performance. Most existing studies fail to consider the partial order relationship between nodes, leading to the generated pseudo-labels not necessarily being representative. This paper proposes a method for constructing a leading tree on graph data to select center nodes. This method integrates the leading tree structure into the GNN and introduces a node similarity-based leading relationship representation layer, which can select the most critical subset of nodes in the graph. These nodes’ pseudo-labels are inferred in a self-supervised manner and added to the node label set for training. Additionally, due to the advantages of the leading tree structure, the number of noisy labels is significantly reduced, greatly alleviating the negative impact of noisy labels on model training. This paper also designs a dual-model pseudo-label training framework, where one model generates pseudo-labels by incorporating leading trees, and the other model is used to predict node labels. Node classification experiments were conducted on six datasets to show the advantages of the proposed architecture.