View-Consistent Heterogeneous Network on Graphs With Few Labeled Nodes.

Area of Science:

• Graph Machine Learning
• Representation Learning

Background:

• Transductive learning on graphs with limited labeled data presents significant challenges due to insufficient supervision.
• Existing methods often rely on single-view self-supervised learning, which may not fully leverage data characteristics.

Purpose of the Study:

• To develop a novel network architecture and training strategy for improved transductive graph learning with very few labels.
• To enhance representation learning by aligning view-agnostic semantics across heterogeneous data views.

Main Methods:

• Proposed a View-Consistent Heterogeneous Network (VCHN) that generates heterogeneous representations for each sample.
• Implemented a view-consistency loss to ensure semantic consistency between different data views.
• Introduced a novel training strategy for reliable pseudolabel generation via mutual supervision between views.

Main Results:

• The VCHN effectively learns better representations by aligning view-agnostic semantics.
• The proposed training strategy enhances VCHN predictions through reliable pseudolabel generation.
• Experimental results on benchmark datasets show superior performance compared to state-of-the-art methods at very low label rates.

Conclusions:

• The VCHN framework offers a powerful approach for transductive graph learning under extreme label scarcity.
• Leveraging multiview representations and mutual supervision significantly improves learning outcomes in low-supervision settings.