Discovering latent node Information by graph attention network.

Area of Science:

• Network science
• Machine learning
• Graph theory

Background:

• Traditional node classification methods are limited to specific graph types.
• Learning effective graph representations is crucial for various network analysis tasks.

Purpose of the Study:

• To introduce Graph Attention based Network Representation (GANR), a novel method for learning graph representations.
• To demonstrate GANR's versatility in handling any given graph structure.

Main Methods:

• Utilizes a graph attention architecture.
• Incorporates graph structure as supervised learning information.
• Learns node representations and attention weights.

Main Results:

• Achieves competitive performance in link prediction, network visualization, and node classification.
• Successfully identifies influential nodes, such as leading venture capital investors and highly cited papers.
• Extracts meaningful attention weights applicable to node centrality measurement.

Conclusions:

• Graph link structures contain latent node information discoverable through appropriate unsupervised learning algorithms like GANR.
• GANR offers a powerful tool for unsupervised representation learning on graphs.
• The attention weights provide interpretable insights into graph importance.