WSSGCN: Wide Sub-stage Graph Convolutional Networks

Chao Wang, Zheng Tang, Hailu Xu

Abstract:

Graph Convolutional Networks (GCNs) have emerged as a potent tool for learning graph representations,
finding applications in a plethora of real-world scenarios. Nevertheless, a significant portion of deep learning
research has predominantly concentrated on enhancing model performance via the construction of deeper
GCNs. Regrettably, the efficacy of training deep GCNs is marred by two fundamental weaknesses: the inade
quacy of conventional methodologies in handling heterogeneous networks, and the exponential surge in model
complexity as network depth increases. This, in turn, imposes constraints on their practical utility. To surmount
these inherent limitations, we propose an innovative approach named the Wide Sub-stage Graph Convolutional
Network (WSSGCN). Our method is an outcome of meticulous observations drawn from classical and graph
convolutional networks, aimed at rectifying the constraints associated with traditional GCNs. Our strategy
involves the conception of a staged convolutional network framework that mirrors the fundamental tenets
of the step-by-step learning process akin to human cognition. This framework prioritizes three distinct forms
of consistency: response-based, feature-based, and relationship-based. Our approach involves three tailored
convolutional networks capturing node/edge, subgraph, and global features. Additionally, we introduce a
novel method to expand graph width for efficient GCN training. Empirical validation on benchmarks highlights
WSSGCN’s superior accuracy and faster training versus conventional GCNs. WSSGCN triumphs over traditional
GCN constraints, significantly enhancing graph representation learning.