Graph Neural Networks with Node Connectivity Strength for Node Classification

As network technologies advance, graph-structured data has grown in importance. Node classification, a key area in analyzing graph data, aims to predict the categories of unlabeled nodes based on graph information. In this paper, we introduce a new Graph Neural Network (GNN) framework called Node Connectivity Strength Graph Neural Network (CSGNN) for node classification. CSGNN uses various powers of the adjacency matrix to capture node connectivity strength, highlighting the role of graph structure. Additionally, we present the Node Connectivity Strength and Similarity Graph Neural Network (CSSGNN) model, which combines node similarity with node connectivity strength to better use node features and graph structure. CSGNN improves traditional GNN aggregation by adding a connectivity strength matrix, while CSSGNN further includes a similarity matrix using cosine similarity to measure node resemblance. In this paper, we also develop a variant model, CSGNN+, which merges the CSGNN framework with a Multilayer Perceptron (MLP). Based on nine benchmark datasets, we conducted experimental tests using 16 baseline models along with CSGNN, CSGNN+, and CSSGNN. The results indicate that CSSGNN demonstrates superior performance. In this paper, we further analyze the datasets features, exploring and verifying how data characteristics affect model performance.