HypBench: Hyperbolic Benchmark for Graph Neural Network Performance

Area of Science:

• Machine Learning
• Graph Theory
• Network Science

Background:

• Graph neural networks (GNNs) show promise in diverse applications like social network analysis and drug discovery.
• GNNs are becoming specialized on benchmark datasets, obscuring performance dependencies on graph properties.
• The relationship between GNN performance and graph topology/features remains an open research question.

Purpose of the Study:

• To introduce HypBench, a comprehensive benchmarking framework for evaluating GNN performance across varied network structures.
• To assess how graph topological and feature properties influence the effectiveness of different GNN architectures.
• To provide insights for selecting appropriate GNN models based on specific data characteristics.

Main Methods:

• Generation of synthetic networks with realistic topological properties and node features using a hyperbolic geometric soft configuration model.
• Systematic evaluation of GNN performance across generated networks with varying properties like topology-feature correlation, degree distributions, local triangle density, and homophily.
• Analysis of the interplay between network structure and node features in determining GNN effectiveness.

Main Results:

• GNN performance is significantly dependent on the intricate interplay between network topology and node feature characteristics.
• The benchmarking framework reveals distinct performance patterns for different GNN architectures under varied graph conditions.
• Identified key graph properties that critically influence GNN model effectiveness across diverse network structures.

Conclusions:

• HypBench offers a versatile tool for evaluating GNNs, facilitating a deeper understanding of their strengths and weaknesses.
• The study underscores the necessity of considering graph-specific properties when selecting or developing GNN models.
• Results guide the development of more robust and adaptable GNN architectures for real-world applications.