Multiview Subgraph Neural Networks: Self-Supervised Learning With Scarce Labeled Data.

Area of Science:

• Machine Learning
• Graph Neural Networks
• Self-Supervised Learning

Background:

• Graph neural networks (GNNs) excel at node classification but require substantial labeled data.
• Low-data regimes hinder GNN performance due to insufficient supervision and overfitting.
• Capturing long-range dependencies is crucial for robust node representation.

Purpose of the Study:

• To develop a novel self-supervised learning (SSL) framework to address the limitations of GNNs in low-data scenarios.
• To enhance node representation by effectively leveraging both local and long-range graph dependencies.
• To improve the classification performance of GNNs on graph data with limited labeled samples.

Main Methods:

• Introduced a self-supervised learning framework named multiview subgraph neural networks (Muse).
• Proposed an information theory-based mechanism to identify subgraphs from input and latent spaces.
• Fused subgraphs capturing local structure and long-range dependencies to create comprehensive node representations.

Main Results:

• Muse effectively augments node representations by capturing both local graph structure and long-range dependencies.
• The framework demonstrates improved performance in node classification tasks within low-data regimes.
• Theoretical generalization error bounds validate the effectiveness of multiview subgraph fusion.

Conclusions:

• Muse offers a promising approach for enhancing GNNs in data-scarce environments.
• Leveraging multiview subgraphs significantly improves the expressiveness of learned node representations.
• The framework provides a robust solution for real-world node classification challenges with limited labeled data.