Homophily-Aware Stratification for Stable GNN Evaluation

Standard random splits in Graph Neural Network evaluation introduce high variance due to inconsistent local neighborhood homophily across folds. This method proposes stratified splits based on relational conditions to ensure test sets reflect the full graph structure, preventing unreliable architecture comparisons.