PSMEKL: Positional and structural multiple empirical kernel learning for node embedding.

Area of Science:

• Graph Neural Networks
• Machine Learning
• Data Science

Background:

• Graph Neural Networks (GNNs) excel at processing graph-structured data.
• Traditional GNNs face challenges in simultaneously capturing both positional and structural node information.
• Existing methods struggle to handle both attributed and non-attributed graphs effectively.

Purpose of the Study:

• To propose a novel method, Positional and Structural Multiple Empirical Kernel Learning (PSMEKL), for enhancing node feature representations.
• To improve the ability of GNNs to capture both positional and structural information.
• To develop a method applicable to both attributed and non-attributed graphs.

Main Methods:

• PSMEKL integrates kernel mapping and community detection techniques.
• It optimizes a dedicated criterion function to generate effective node embeddings.
• The method captures both positional relationships and global graph structures.

Main Results:

• PSMEKL enhances the distinction between nodes with different structures.
• It ensures nodes with similar connectivity patterns share similar embedding features.
• Preprocessing node features with PSMEKL significantly improves GNN performance across extensive experiments.

Conclusions:

• PSMEKL offers a robust approach to node embedding for GNNs.
• The method effectively addresses limitations of traditional GNNs in capturing multi-faceted node information.
• PSMEKL demonstrates broad applicability and performance enhancement for various graph types.