Are Graph Convolutional Networks With Random Weights Feasible?

Area of Science:

• Graph Neural Networks
• Machine Learning
• Network Analysis

Background:

• Graph Convolutional Networks (GCNs) excel at learning node representations but face scalability challenges with large graphs due to computationally intensive training.
• Existing GCN extensions focus on sampling and feature aggregation, yet learning efficiency remains a bottleneck for massive datasets.
• Training vanilla GCNs requires full dataset batch gradient descent, hindering application on large-scale graph data.

Purpose of the Study:

• To investigate the feasibility of using random weights to enhance GCN training efficiency.
• To propose and theoretically analyze a novel GCN model incorporating random weights.
• To empirically validate the effectiveness and efficiency of the proposed model on benchmark datasets.

Main Methods:

• Introduced Graph Convolutional Networks with Random Weights (GCN-RW) by modifying convolutional layers with random filters.
• Adjusted the learning objective by incorporating a regularized least squares loss function.
• Conducted theoretical analyses on approximation upper bound, structure complexity, stability, and generalization, supported by mathematical proofs.

Main Results:

• GCN-RW demonstrated effectiveness and efficiency in semi-supervised node classification tasks across multiple benchmark datasets.
• The proposed model achieved accuracy comparable to or better than state-of-the-art approaches.
• GCN-RW significantly reduced training time compared to existing methods.

Conclusions:

• GCN-RW offers a promising solution for efficient large-scale graph representation learning.
• The model provides a strong balance between predictive accuracy and computational efficiency.
• Random weights present a viable strategy for accelerating GCN training without sacrificing performance.