Graph-Graph Similarity Network.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Network Science

Background:

• Graph learning aims to predict labels for entire graphs.
• Graph Neural Networks (GNNs) learn graph embeddings but overlook inter-graph relationships.
• Existing methods struggle to capture the broader context of relationships among graphs.

Purpose of the Study:

• To propose a novel graph-graph (G2G) similarity network for enhanced graph learning.
• To address the limitation of GNNs ignoring relationships among graphs.
• To transform graph learning into a node label propagation problem on a SuperGraph.

Main Methods:

• Constructing a SuperGraph where nodes represent input graphs and edge weights signify graph similarity.
• Utilizing an adversarial autoencoder to align graph embeddings to a prior distribution.
• Designing the G2G similarity network to learn inter-graph similarities, forming the SuperGraph's adjacency matrix.

Main Results:

• The G2G similarity network effectively captures relationships among graphs.
• Graph learning is successfully reframed as a node label propagation task on the SuperGraph.
• Experiments demonstrate significant effectiveness on multiple classification and regression benchmarks.

Conclusions:

• The proposed G2G similarity network enhances graph label prediction by considering inter-graph relationships.
• This approach offers a powerful new framework for graph learning tasks.
• The method shows broad applicability and effectiveness across diverse graph datasets.