MBGCN: Multi-View Block-Wise Graph Convolutional Networks on Large-Scale Graphs.

Area of Science:

• Graph Neural Networks
• Machine Learning
• Computer Science

Background:

• Traditional graph convolutional networks (GCNs) face scalability issues with large graphs.
• Existing methods like edge sparsification and node sampling lead to information loss and bias.
• Multi-view fusion techniques struggle to balance inter-view consistency and intra-view diversity.

Purpose of the Study:

• To propose a multi-view block-wise graph convolutional network (MB-GCN) for large-scale graph analysis.
• To address the computational inefficiency and information loss in existing GCN methods.
• To improve performance in multi-view semi-supervised classification tasks.

Main Methods:

• Implemented a node segmentation module to partition nodes into view-specific subsets, reducing complexity.
• Employed alternating graph convolution and graph structure learning within blocks using a shared-weight strategy for enhanced feature extraction.
• Introduced a global fusion module with a cross-view inter-block loss for aligning representations and alleviating over-smoothing.

Main Results:

• The proposed MB-GCN significantly outperforms state-of-the-art methods on diverse large-scale graph datasets.
• Demonstrated superior scalability and memory efficiency compared to existing approaches.
• Achieved improved results in multi-view semi-supervised classification tasks.

Conclusions:

• The MB-GCN effectively handles large-scale graphs by reducing computational complexity while preserving local information.
• The method successfully leverages multi-view information, achieving better inter-view consistency and intra-view diversity.
• MB-GCN offers a scalable and efficient solution for multi-view semi-supervised graph classification.