Finding core labels for maximizing generalization of graph neural networks.

Area of Science:

• Machine Learning
• Graph Representation Learning

Background:

• Graph neural networks (GNNs) are widely used for semi-supervised graph representation learning.
• Research has primarily focused on GNN methodologies, with less emphasis on training data selection.
• High-quality training data is crucial for effective semi-supervised learning.

Purpose of the Study:

• To address the problem of training data selection for GNNs in node classification.
• To identify a representative subset of nodes that optimizes GNN performance.
• To propose and validate a new hypothesis for selecting optimal training data in graph learning.

Main Methods:

• Introduction and elaboration of the 'subset hypothesis' for graph data, analogous to the lottery ticket hypothesis.
• Development of an efficient algorithm to identify the core data subset within a graph.
• Experimental validation across various datasets and GNN architectures.

Main Results:

• The proposed subset hypothesis posits the existence of a core data subset that captures essential dataset properties.
• The developed algorithm effectively identifies this core subset for GNN training.
• GNNs trained on the selected core subset consistently demonstrated performance improvements.

Conclusions:

• The quality and selection of training data significantly impact GNN performance.
• The subset hypothesis provides a novel framework for understanding optimal data selection in graph learning.
• The proposed method offers an efficient approach to enhance GNNs through strategic data selection.