GPENs: Graph Data Learning With Graph Propagation-Embedding Networks.

Area of Science:

• Machine Learning
• Pattern Recognition
• Graph Theory

Background:

• Compact graph data representation is crucial for machine learning.
• Graph Neural Networks (GNNs) are increasingly used for graph-structured data.
• Existing methods may lack efficiency or unified approaches.

Purpose of the Study:

• Introduce Graph Propagation-Embedding Networks (GPENs) for effective graph representation.
• Unify traditional graph propagation with deep embedding techniques.
• Develop a computationally efficient and explainable graph learning model.

Main Methods:

• Developed a novel propagation-embedding architecture integrating feature propagation and low-dimensional embedding.
• Utilized exact and approximate formulations with closed-form solutions for efficiency.
• Extended the model to Multiple GPENs (M-GPENs) for multi-graph data.

Main Results:

• GPENs demonstrate effectiveness in semi-supervised learning tasks.
• The model offers a well-motivated and explainable approach.
• Achieved compactivity and computational efficiency through closed-form solutions.

Conclusions:

• GPENs provide a powerful and efficient method for graph-structured data representation and learning.
• The unified architecture and efficient solutions offer significant advantages.
• M-GPENs extend applicability to complex datasets with multiple graph structures.