Property-Driven Evaluation of GNN Expressiveness at Scale: Datasets, Framework, and Study

Researchers developed a comprehensive evaluation framework for Graph Neural Networks (GNNs) using formal specification methods, creating 336 new datasets to test GNN expressiveness across 16 fundamental graph properties. The study reveals that no single pooling approach consistently performs well across all properties, with attention-based pooling excelling in generalization while second-order pooling provides better sensitivity.

• →New evaluation methodology introduces 336 datasets with over 10,000 labeled graphs each to test GNN expressiveness systematically.
• →Study covers 16 fundamental graph properties critical to distributed systems, knowledge graphs, and biological networks.
• →Attention-based pooling methods excel in generalization and robustness compared to other approaches.
• →Second-order pooling provides superior sensitivity but lacks consistency across different graph properties.
• →Research highlights fundamental limitations in current GNN architectures and suggests adaptive property-aware pooling as future direction.