NCGNN: Node-Level Capsule Graph Neural Network for Semisupervised Classification.

Area of Science:

• Graph Neural Networks
• Machine Learning
• Network Science

Background:

• Existing graph neural networks (GNNs) use simple aggregation (sum/average) for message passing, leading to interpretability issues and feature overmixing.
• Feature overmixing causes oversmoothing, hindering long-range dependency capture and performance on graphs with low homophily or heterophily.

Purpose of the Study:

• To propose a novel Node-Capsule Graph Neural Network (NCGNN) with an improved message passing scheme.
• To address the limitations of interpretability and feature overmixing in current GNNs.
• To enhance node representation learning for graphs exhibiting homophily or heterophily.

Main Methods:

• NCGNN represents nodes using node-level capsules, with each capsule extracting distinctive node features.
• A dynamic routing procedure adaptively selects relevant capsules from a subgraph identified by a graph filter.
• This selective aggregation prevents overmixing, restraining irrelevant messages and preserving important features.

Main Results:

• NCGNN effectively mitigates the oversmoothing issue, enabling better capture of long-range dependencies.
• The model demonstrates improved node representation learning on graphs with both homophily and heterophily.
• Experiments show NCGNN outperforms state-of-the-art methods in semisupervised node classification tasks.

Conclusions:

• NCGNN offers an interpretable message passing scheme, identifying significant features without post hoc analysis.
• The proposed method enhances GNN performance by selectively aggregating informative node capsules.
• NCGNN provides superior node classification accuracy on diverse graph structures, including heterophilic ones.