WSSGCN: Wide Sub-stage Graph Convolutional Networks

Abstract

Graph Convolutional Networks (GCNs) have become a powerful tool for learning graph representations, with applications in many real-world scenarios. However, much research has focused on improving performance by building deeper GCNs, which face two key issues: difficulty in handling heterogeneous networks and a sharp increase in model complexity with greater depth, limiting practical use. To address these challenges, we propose the Wide Sub-stage Graph Convolutional Network (WSSGCN), inspired by both classical and graph convolutional networks. Our approach introduces a staged framework, reflecting a step-by-step learning process, and emphasizes three types of consistency: response-based, feature-based, and relationship-based. It employs tailored networks to capture node/edge, subgraph, and global features, alongside a method to increase graph width for efficient training. Benchmarks show WSSGCN achieves higher accuracy and faster training than conventional GCNs, overcoming their limitations and significantly improving graph representation learning.