Bayesian graph convolutional network with partial observations.

Area of Science:

• Machine Learning
• Graph Neural Networks
• Bayesian Inference

Background:

• Graph convolutional networks (GCNs) excel at non-grid data but typically require fully observed node features.
• Many real-world applications lack complete node features, posing a challenge for existing GCNs.

Purpose of the Study:

• To propose a novel Bayesian framework for graph node classification that does not rely on node features.
• To develop a GCN model capable of handling graphs with missing or unavailable node features.

Main Methods:

• Introduced pseudo-features generated via a stochastic process for each graph node.
• Incorporated a hidden space structure preservation term to maintain data distribution consistency.
• Employed variational inference for efficient training and prediction algorithms.

Main Results:

• The proposed GCN model significantly outperforms traditional methods in graph node classification tasks.
• Achieved an average performance improvement of 9% across various datasets.
• Demonstrated the effectiveness of the Bayesian approach in feature-less graph scenarios.

Conclusions:

• The novel Bayesian graph convolutional network effectively performs node classification without node features.
• The method offers a significant advancement for applications with incomplete graph data.
• The proposed approach provides a robust and efficient solution for feature-less graph analysis.