Toward Robust Graph Semi-Supervised Learning Against Extreme Data Scarcity.

Area of Science:

• Machine Learning
• Data Science
• Graph Analytics

Background:

• Graph Neural Networks (GNNs) require extensive human-annotated data for web mining.
• Semi-supervised learning on graphs is challenging with few labeled nodes due to difficulties in feature-label propagation and decision boundary learning.

Purpose of the Study:

• To develop a robust graph predictive model for low-data scenarios.
• To address limitations of self-training on graph-structured data by capturing structural and semantic knowledge.

Main Methods:

• Proposed a novel graph data augmentation framework named Augmented Graph Self-Training (AGST).
• AGST incorporates two new augmentation modules: structural and semantic.
• Utilized a decoupled Graph Self-Training (GST) backbone.

Main Results:

• Comprehensive evaluations were conducted on semi-supervised node classification tasks.
• The framework was tested under various limited labeled-node data scenarios.
• Experimental results demonstrated the framework's effectiveness in low-data contexts.

Conclusions:

• The proposed AGST framework significantly enhances the robustness of graph predictive models with limited labeled data.
• The novel data augmentation modules contribute uniquely to node classification performance.
• AGST offers a promising solution for sustainable graph semi-supervised learning.