Multi-view graph pooling with coarsened graph disentanglement.

Area of Science:

• Graph Neural Networks
• Machine Learning
• Information Theory

Background:

• Multi-view graph pooling enhances graph-level tasks by integrating information from multiple perspectives.
• Existing methods lack explicit control over pooling and theoretical understanding of view relationships.

Purpose of the Study:

• To introduce GDMGP, a novel multi-view graph pooling method based on graph disentanglement and information theory.
• To improve graph representation by creating a structured, disentangled coarsened graph.

Main Methods:

• A novel view mapper integrates node and topology information.
• A conditional entropy-based fusion mechanism regulates task-relevant information.
• Mutual information minimization disentangles fused views into task-relevant and irrelevant subgraphs.

Main Results:

• Theoretically demonstrated that GDMGP's coarsened graph outperforms any single input view.
• Experimental validation on seven public datasets confirmed GDMGP's effectiveness.
• GDMGP enhances clarity and utility of graph representations.

Conclusions:

• GDMGP offers a principled approach to multi-view graph pooling.
• The method provides superior graph representations for downstream tasks.
• GDMGP advances graph representation learning through disentanglement and information theory.