Self-Supervised Node Classification with Strategy and Actively Selected Labeled Set.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Self-supervised learning (SSL) enhances graph neural networks (GNNs) for classification with limited labeled data by leveraging unlabeled nodes.
• Existing SSL pretext tasks lack universal optimality across diverse datasets and require careful hyperparameter tuning.

Purpose of the Study:

• To develop an automated framework for selecting optimal self-supervised pretext tasks and hyperparameters for GNNs.
• To improve semi-supervised node classification accuracy by integrating automated SSL with active learning.

Main Methods:

• Proposed a novel auto graph self-supervised learning framework.
• Integrated a one-shot active learning method for enhanced label selection.
• Evaluated performance on three real-world citation datasets.

Main Results:

• Automatically optimized pretext tasks achieved comparable or superior GNN classification accuracy to manual methods.
• Active selection of labeled nodes further boosted classification performance compared to random selection.
• Both automated optimization and active selection significantly contributed to semi-supervised node classification.

Conclusions:

• The proposed framework automates the selection and optimization of SSL pretext tasks for GNNs.
• Active learning enhances the efficiency and effectiveness of semi-supervised node classification.
• This approach reduces the need for expert experience in configuring GNNs for limited-label scenarios.