Semantic Graph Neural Network with Multi-measure Learning for Semi-supervised Classification

TL;DR

This paper addresses the brittleness of GNNs to graph structure by learning task-specific graph representations that capture both local and global structure. It introduces ML-SGNN, which jointly learns a feature graph via multi-measure learning, a topology graph, and a semantic graph encoding global context, then fuses them through attention-guided aggregation. The model leverages supervised loss and two $l_{2,1}$-based unsupervised regularizers to encourage cross-embedding consistency, achieving state-of-the-art results on six real-world datasets and demonstrating robustness across label rates. The approach offers scalable, adaptive graph structure learning with practical implications for robust semi-supervised node classification in noisy or incomplete graphs.

Abstract

Graph Neural Networks (GNNs) have attracted increasing attention in recent years and have achieved excellent performance in semi-supervised node classification tasks. The success of most GNNs relies on one fundamental assumption, i.e., the original graph structure data is available. However, recent studies have shown that GNNs are vulnerable to the complex underlying structure of the graph, making it necessary to learn comprehensive and robust graph structures for downstream tasks, rather than relying only on the raw graph structure. In light of this, we seek to learn optimal graph structures for downstream tasks and propose a novel framework for semi-supervised classification. Specifically, based on the structural context information of graph and node representations, we encode the complex interactions in semantics and generate semantic graphs to preserve the global structure. Moreover, we develop a novel multi-measure attention layer to optimize the similarity rather than prescribing it a priori, so that the similarity can be adaptively evaluated by integrating measures. These graphs are fused and optimized together with GNN towards semi-supervised classification objective. Extensive experiments and ablation studies on six real-world datasets clearly demonstrate the effectiveness of our proposed model and the contribution of each component.

Figures (9)

• Figure 1: The framework of ML-SGNN model, consists of three GNN modules designed for node embedding, two attention layers used for graph fusion, and an optimizer used for improve generalization, where feature subgraphs are constructed by multi-measures and semantic graph is constructed by optimized semantic encoding to preserve the local and global structure. More details are described in Section \ref{['sec3']}.
• Figure 2: Performance of ML-SGNN and its variants.
• Figure 3: Comparision of ML-SGNN with average-measure and single-measure.
• Figure 4: Heatmap of p-values obtained from Wilcoxon test.
• Figure 5: Critical Difference (CD) diagram from Friedman test.