Leveraging similarity and structural correlation for attention-based graph embedding

Graph attention networks (GATs) are a prevalent method for graph embedding, utilising attention mechanisms to aggregate first-order neighbourhood node information. However, their inability to adequately consider high-order neighbourhood nodes and structural information poses limitations. To address this gap, we propose a novel joint attention graph embedding model integrating similar networks and structural correlation. By introducing the concept of structural correlation, our model comprehensively incorporates both node content features and joint node topological structure features when computing attention scores. Experimental results showcase the efficacy of our approach, yielding significant accuracy improvements of 2.70%, 3.94%, and 2.60% on the Cora, Citeseer, and Pubmed datasets, respectively, compared to traditional GATs. Our proposed method significantly enhances node embedding representation, underscoring its importance in improving performance in node classification tasks.