Semisupervised Graph Neural Networks for Graph Classification.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Graph classification is crucial for analyzing graph-structured data across various applications.
• Graph neural networks (GNNs) excel in supervised graph classification but struggle with limited labeled data.
• Existing semisupervised GNNs are primarily designed for node classification, not graph classification.

Purpose of the Study:

• To propose a general semisupervised GNN framework for graph classification tasks.
• To address the challenge of insufficient labeled data in practical graph classification scenarios.
• To leverage both limited labeled and abundant unlabeled graph data for improved classification.

Main Methods:

• Developed a novel semisupervised GNN framework utilizing two complementary GNNs.
• Employed collaborative learning between GNNs on both labeled and unlabeled graphs.
• Incorporated pseudo-labeling of high-confidence graph examples to augment the training dataset.

Main Results:

• Demonstrated the effectiveness of the proposed framework on benchmark datasets.
• Showcased successful application in scenarios with few and extremely few labeled graphs.
• Achieved high-quality classifier performance through collaborative learning and pseudo-labeling.

Conclusions:

• The proposed semisupervised GNN framework significantly enhances graph classification performance.
• The framework effectively utilizes limited labeled data alongside abundant unlabeled data.
• This approach offers a robust solution for practical graph classification challenges.