MGAT: Multi-view Graph Attention Networks.

Area of Science:

• Graph embedding
• Network science
• Machine learning

Background:

• Existing graph embedding methods primarily focus on single-view networks, limiting their ability to capture complex relationships.
• Real-world systems often exhibit multiple, interconnected types of relationships among entities.
• There is a need for advanced methods to effectively model these multi-view networks.

Purpose of the Study:

• To propose a novel approach for multi-view graph embedding named Multi-view Graph Attention Networks (MGAT).
• To develop an attention-based architecture for learning node representations within each view.
• To introduce a method for collaboratively integrating information from multiple views.

Main Methods:

• Developed an attention-based architecture for single-view node representation learning.
• Incorporated a novel regularization term to constrain network parameters.
• Employed a view-focused attention mechanism to aggregate view-wise representations.

Main Results:

• The proposed Multi-view Graph Attention Networks (MGAT) were evaluated on several real-world datasets.
• MGAT demonstrated superior performance compared to existing state-of-the-art baseline methods.
• The approach effectively captures and integrates multiple relationship types.

Conclusions:

• MGAT offers a powerful solution for multi-view graph embedding.
• The attention-based architecture and view-focused aggregation are key to its success.
• This method advances the ability to model complex systems with diverse relationships.