Trainable-parameter-free structural-diversity message passing for graph neural networks.

Area of Science:

• Graph Neural Networks
• Machine Learning
• Network Science

Background:

• Graph Neural Networks (GNNs) excel at structured data but struggle with heterogeneous neighborhoods and complex features.
• Mainstream GNNs often homogenize representations due to uniform neighbor aggregation and many learnable parameters.
• This limits adaptability in low-supervision or imbalanced datasets, leading to semantic degradation.

Purpose of the Study:

• Introduce a parameter-free GNN framework, Structural-Diversity Graph Neural Network (SDGNN), to address representation homogenization.
• Operationalize structural diversity in message passing to better model heterogeneous graph neighborhoods.
• Enhance adaptability across diverse graph structures and challenging learning conditions.

Main Methods:

• Propose Structural-Diversity Message Passing (SDMP) with within-group statistics and cross-group selection.
• Incorporate structure-driven and feature-driven partitioning strategies.
• Utilize a normalized-propagation-based global structural enhancer for improved adaptability.

Main Results:

• SDGNN consistently outperforms mainstream GNNs across nine benchmark datasets and a PubMed citation network.
• Demonstrates superior performance under low supervision and class imbalance conditions.
• Shows enhanced adaptability in cross-domain transfer learning tasks.

Conclusions:

• SDGNN effectively models structural diversity in graphs, overcoming limitations of existing GNNs.
• The parameter-free design and novel message-passing mechanism offer improved representation learning.
• SDGNN presents a robust and adaptable solution for real-world graph data challenges.