Scalable and Effective Temporal Graph Representation Learning With Hyperbolic Geometry.

Area of Science:

• Graph Neural Networks
• Dynamic Systems
• Hyperbolic Geometry

Background:

• Real-world graphs exhibit complex, continuous-time dynamics.
• Existing Euclidean-based temporal graph neural networks (TGNNs) struggle with hierarchical structures, leading to high-distortion embeddings.
• Euclidean space's limitations hinder the accurate representation of complex graph topologies.

Purpose of the Study:

• To propose a scalable and effective TGNN using hyperbolic geometries for continuous-time dynamic graph (CTDG) representation.
• To enhance the representation capabilities of TGNNs by overcoming Euclidean limitations.
• To develop a unified framework that captures evolving behaviors and hierarchical structures simultaneously.

Main Methods:

• Introduced a scalable TGNN with hyperbolic geometries (STGNh).
• Integrated a memory-based module (hyperbolic update gate - HuG) for efficient temporal dynamics storage.
• Developed a structure-based module (hyperbolic temporal Transformer - HyT) for complex structure capture and node embedding generation.

Main Results:

• STGNh demonstrates scalability to billion-scale graphs.
• The model effectively captures both evolving behaviors and hierarchical graph structures.
• Extensive experiments show STGNh significantly outperforms baseline methods on various downstream tasks.

Conclusions:

• Hyperbolic geometry offers enhanced representation capabilities for complex CTDGs compared to Euclidean geometry.
• The proposed STGNh framework provides a powerful and scalable solution for modeling dynamic graph data.
• This approach significantly improves performance in downstream tasks involving complex temporal graph structures.