Graph Few-Shot Learning via Restructuring Task Graph.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Graph few-shot learning (FSL) methods face challenges with isolated nodes and insufficient task information.
• Existing models struggle with poor node embeddings and limited expressive ability.

Purpose of the Study:

• To propose a novel metric-based graph few-shot learning approach via restructuring task graphs (GFL-RTG).
• To enhance node embeddings and model expressiveness in graph FSL.

Main Methods:

• Restructure task graphs by adding class nodes and a task node.
• Utilize a graph pooling network to learn a task embedding (task node).
• Input the restructured graph into a metric-based graph neural network (GNN) for few-shot learning.

Main Results:

• The proposed GFL-RTG method addresses deficiencies in node embeddings caused by isolated nodes.
• The integration of task information enhances the model's expressive ability.
• Experiments show the method generally outperforms state-of-the-art baselines on graph-structured datasets.

Conclusions:

• The GFL-RTG approach offers an effective solution for graph few-shot learning.
• Restructuring task graphs improves performance in few-shot learning scenarios.
• The method demonstrates superior performance compared to existing graph FSL techniques.