Enhancing Semi-Supervised Learning in Graph Convolutional Networks: The AEWGS Framework

In recent years, graph convolutional networks (GCNs) have been widely used in graph analysis tasks, especially for semi-supervised learning. However, GCNs and their variants usually rely on unidirectional information propagation, requiring all nodes to participate in the training process. This leads to performance degradation when dealing with large datasets. To address these challenges, we propose a novel end-to-end framework, AEWGS. AEWGS starts with an edge sampling algorithm for initial sample selection, and then uses a degree-based labeling algorithm to determine the node weights by combining the label information with the node degree. Next, a K-means clustering method divides the weighted graph into clusters based on the node weights, generating optimized subgraph sets to improve training efficiency and detail retention. Finally, AEWGS accelerates convergence by combining the asymptotic approach with a parameter sharing strategy for incremental GCN training. Experiments on four public datasets show that AEWGS outperforms eleven GCN variants in semi-supervised node classification, improving classification accuracy and training efficiency.