The Post-GCN Decade Revisited: Curvature-Stratified Evaluation of Relational Learning
Researchers have introduced a new framework for evaluating relational learning models, moving beyond standard leaderboards that average performance across diverse datasets. This new approach stratifies datasets by their geometric properties, revealing that model performance is highly dependent on these intrinsic geometries. The study evaluated 18 models, including GCNs and GFMs, across 14 datasets, finding that rankings shift significantly across different curvature regimes, suggesting that some advanced models may offer diminishing returns in specific geometric contexts. AI
IMPACT Introduces a more nuanced evaluation method that could lead to more robust and interpretable comparisons of future relational learning models.