Graph contrastive learning with virtual nodes for few-shot semi-supervised classification.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Contrastive learning shows promise for graph few-shot semi-supervised learning (GFSSL).
• Existing GFSSL methods struggle with insufficient supervision and imbalanced positive samples, leading to representation collapse.
• Representation collapse occurs when node embeddings overly specialize in specific subclass features.

Purpose of the Study:

• To propose a novel GFSSL framework addressing representation collapse.
• To enhance supervisory signals and balance contrastive pairs using label-guided virtual nodes.
• To improve the performance of GFSSL by preventing node embeddings from collapsing into single subclasses.

Main Methods:

• A graph encoder is pre-trained to mine high-confidence pseudo-labels from local neighborhoods.
• Label-guided virtual nodes are introduced as class-level proxies to aggregate features across subclasses.
• Virtual nodes are integrated into the graph to connect semantically similar nodes and expand positive samples in contrastive learning.

Main Results:

• The proposed model significantly outperforms the best baseline methods across six datasets.
• An average performance improvement of 2.06% was achieved.
• The method effectively prevents representation collapse by providing richer subclass features.

Conclusions:

• The novel GFSSL framework with label-guided virtual nodes effectively addresses limitations of existing methods.
• The approach enhances representation learning by supplementing supervision and balancing contrastive pairs.
• This work offers a promising direction for improving few-shot learning on graph data.