Domain-adaptive message passing graph neural network.

Area of Science:

• Machine Learning
• Graph Neural Networks
• Network Science

Background:

• Cross-network node classification (CNNC) addresses label scarcity in target networks by leveraging data from source networks.
• Existing methods struggle with effective knowledge transfer across different network domains.

Purpose of the Study:

• To develop a novel domain-adaptive graph neural network (DM-GNN) for improved CNNC.
• To enhance the transferability of node representations across networks.

Main Methods:

• A GNN encoder with dual feature extractors for ego and neighbor embeddings.
• A label propagation classifier incorporating neighborhood information.
• A label-aware propagation scheme for source networks.
• Conditional adversarial domain adaptation considering neighborhood-refined labels.

Main Results:

• The proposed DM-GNN effectively learns informative and transferable node representations.
• The method demonstrates superior performance compared to eleven state-of-the-art CNNC techniques.
• Conditional adversarial domain adaptation improves cross-network distribution matching.

Conclusions:

• DM-GNN offers a robust solution for cross-network node classification.
• Integrating GNNs with domain adaptation techniques enhances knowledge transfer capabilities.
• The proposed methods significantly advance the field of CNNC.