Hyperbolic Graph Convolutional Neural Networks.

Area of Science:

• Machine Learning
• Graph Neural Networks
• Hyperbolic Geometry

Background:

• Graph convolutional neural networks (GCNs) embed graph nodes into Euclidean space, causing significant distortion for scale-free or hierarchical graphs.
• Hyperbolic geometry offers a promising alternative for embeddings with reduced distortion.

Purpose of the Study:

• To propose the first inductive hyperbolic GCN (HGCN) for learning node representations in hierarchical and scale-free graphs.
• To address challenges in defining neural network operations and mapping Euclidean features in hyperbolic space.

Main Methods:

• Deriving GCN operations within the hyperboloid model of hyperbolic space.
• Mapping Euclidean input features to hyperbolic embeddings with trainable curvature at each layer.
• Developing HGCN as an inductive model for graph representation learning.

Main Results:

• HGCN effectively preserves the hierarchical structure of graphs.
• Achieved up to 63.1% error reduction in ROC AUC for link prediction.
• Achieved up to 47.5% F1 score improvement for node classification compared to Euclidean GCNs.
• Demonstrated state-of-the-art performance on the Pubmed dataset, even with low-dimensional embeddings.

Conclusions:

• HGCN successfully combines GCN expressiveness with hyperbolic geometry for superior node representation learning.
• The proposed method offers significant performance gains over existing GCNs on complex graph structures.
• HGCN provides a powerful new tool for analyzing and understanding hierarchical and scale-free networks.