IGCN: A Provably Informative GCN Embedding for Semi-Supervised Learning With Extremely Limited Labels

Area of Science:

• Graph Neural Networks
• Representation Learning
• Machine Learning

Background:

• Graph Neural Networks (GNNs) excel at representation learning for graph-structured data.
• Limited labels in graph data often cause overfitting and poor model performance.

Purpose of the Study:

• Propose Informative Graph Convolutional Network (IGCN) to enhance GNNs with limited labels.
• Obtain informative embeddings by discarding task-irrelevant graph information using mutual information.

Main Methods:

• Optimize IGCN using a surrogate objective due to intractable mutual information computation for irregular data.
• Utilize semi-supervised classification and prototype-based supervised contrastive learning for the first objective term.
• Employ a graph encoder-decoder module and GCN_Info architecture to minimize reconstruction loss for the second objective term, preserving initial embedding information.

Main Results:

• The proposed GCN_Info architecture provably alleviates information loss.
• IGCN effectively preserves useful information from initial embeddings.
• Experimental results demonstrate IGCN's superior performance over state-of-the-art methods on 7 datasets.

Conclusions:

• IGCN framework successfully tackles the overfitting problem in GNNs with limited labels.
• The method enhances representation learning by focusing on informative embeddings.
• IGCN represents a significant advancement in graph representation learning.