Performance Enhancement Strategies for Node Classification Based on Graph Community Structure Recognition

For current Transformers, although current learnable graph community structures help enhance the information capture of low-degree nodes, these approaches primarily aim to improve the Transformer model rather than deeply exploring the underlying relationships between nodes. As a result, they face limitations in enhancing the role of low-degree nodes within the community. To address these limitations, this paper introduces two innovative methods: Spectral-Mean Synergistic Clustering (SMSC) and the Semantic Proximity Evaluation Mechanism (SPEM). SMSC improves the construction of graph communities, enabling more accurate classification of nodes into different communities. SPEM evaluates the potential value and similarity of low-degree nodes within the same community to establish better connections. The proposed algorithm demonstrates significant performance improvements in node classification. Compared to the baseline model, it achieves a 1.64% improvement on the Cora dataset and a 4.05% improvement on the much larger WikiCS dataset. Experimental results verify that combining spectral-mean co-optimization with semantic proximity assessment significantly enhances the processing of low-degree node information, especially in larger-scale datasets.