Improving Node Classification through Convolutional Networks Built on Enhanced Message-Passing Graph.

Area of Science:

• Graph Neural Networks
• Machine Learning
• Data Science

Background:

• Graph Convolutional Networks (GCNs) struggle with effective message propagation in sparse graphs due to over-smoothing and overfitting with limited labeled data.
• Enhancing message propagation is crucial for accurate node classification in graphs with scarce labels.

Purpose of the Study:

• To address the limitations of GCNs in sparse graphs by introducing an Enhanced Message-Passing Graph (EMPG) framework.
• To improve node classification accuracy by enabling more efficient message propagation to distant nodes.

Main Methods:

• Node mapping to a latent space via graph embedding to capture structural information.
• Construction of EMPG by adding connections between nodes in proximity in the latent space.
• Utilizing dropout on EMPG to extract graph variants and building multichannel GCNs for aggregated predictions.

Main Results:

• The proposed method effectively avoids over-smoothing by reducing the need for deep GCN layers.
• The EMPG framework demonstrates robustness and efficiency in node classification tasks.
• Experimental results show significant improvements in node classification accuracy compared to existing methods.

Conclusions:

• The EMPG framework offers a flexible and effective solution for node classification in sparse graphs.
• The method successfully mitigates GCN limitations like over-smoothing and overfitting.
• This approach provides a robust and efficient means to enhance message propagation for improved graph-based learning.