Multi-graph Fusion Graph Convolutional Networks with pseudo-label supervision.

Area of Science:

• Machine Learning
• Graph Neural Networks
• Data Mining

Background:

• Graph convolutional networks (GCNs) excel at learning from graph data.
• Current methods struggle to effectively fuse topological and feature information for classification.
• Real-world graphs often contain noisy or incomplete edges, impacting GCN performance.

Purpose of the Study:

• To develop a more robust and accurate graph structure learning method.
• To enhance semi-supervised node classification by effectively integrating multi-graph and node features.
• To address the challenge of missing labels in semi-supervised learning scenarios.

Main Methods:

• Proposed Multi-graph Fusion Graph Convolutional Networks (MFGCN).
• Learned a connected embedding by fusing multi-graphs and node features.
• Implemented a pseudo-label generation mechanism based on node feature similarity to handle missing labels.

Main Results:

• MFGCN demonstrated superior performance in semi-supervised node classification tasks.
• The proposed method effectively propagates node features over multi-graphs for improved embeddings.
• Pseudo-label supervision enhanced classification accuracy, especially with limited labeled data.

Conclusions:

• MFGCN offers a robust approach to graph structure learning by adaptively fusing diverse graph information.
• The method significantly improves semi-supervised node classification accuracy compared to state-of-the-art techniques.
• The pseudo-labeling strategy effectively mitigates issues arising from missing labels in graph datasets.