Multi-Subgraph Fusion: An Innovative Approach for Block Matrix Graph Convolutional Networks

Abstract Graph Convolutional Networks (GCNs) is a dominant approach for graph representation learning through neighborhood aggregation.However, existing GCN methods rely on a single structural view that only captures direct neighborhood connections. This limitation overlooks important long-range dependencies and global topological patterns, leading to suboptimal node representations for downstream tasks. To address these limitations, we propose a Multi-Block Graph Convolutional Network (MBGCN) that constructs two complementary subgraphs of the graph: a graph diffusion-based subgraph that captures global topological context through information propagation, and a \(K\) -nearest neighbor subgraph that preserves fine-grained local structural patterns. Additionally, the original graph structure is preserved as a structural subgraph to retain the graph’s inherent connectivity information. These three subgraphs work collaboratively to encode both macro-level community structures and micro-level neighborhood details, enabling more comprehensive node representations. Specifically, MBGCN consists of three key components: (1) A subgraph generation module that leverages graph diffusion and $K$-nearest neighbor strategies to generate complementary subgraphs. (2) The block modeling module allows the model to assign weights to nodes of different classes, considering the homogeneity and heterogeneity characteristics of the network. (3) A block-guided graph convolution module that adaptively integrates the learned representations from different views to generate comprehensive node embeddings. Extensive experiments on three real-world datasets demonstrate that MBGCN outperforms eight state-of-the-art baselines. Notably, it achieves an improvement of 5.29% over the second-best method on the Texas dataset.