LGLNN: Label Guided Graph Learning-Neural Network for few-shot learning.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Vision

Background:

• Graph Neural Networks (GNNs) are used for few-shot learning (FSL).
• Existing GNN-based FSL methods often overlook pairwise label constraints and independent edge learning, limiting optimal graph construction.
• This leads to suboptimal performance in few-shot classification tasks.

Purpose of the Study:

• To propose a novel Label Guided Graph Learning-Neural network (LGLNN) for few-shot learning tasks.
• To address limitations in existing GNN-based FSL by incorporating label information for optimal graph metric learning.
• To enable cooperative and consistent metric learning across all graph edges.

Main Methods:

• The proposed Label Guided Graph Learning-Neural network (LGLNN) model is introduced.
• LGLNN incorporates label information to learn an optimal metric graph for GNNs.
• Pairwise constraint propagation is employed, and metric learning for edges aggregates information from neighboring edges.

Main Results:

• Experimental results demonstrate the effectiveness of the proposed LGLNN method.
• The LGLNN model achieves better performance compared to existing approaches.
• The model successfully learns metrics for graph edges cooperatively and consistently.

Conclusions:

• The LGLNN model offers an effective approach to few-shot learning using GNNs.
• Incorporating label constraints and consistent edge learning significantly improves performance.
• LGLNN provides a robust framework for metric learning in graph-based FSL.