How Much Does Message Passing Matter? A Drop-In Study of GNN Layers for Neural Network Graph Regression
The study tested ten different Message Passing (MP) layers across four existing Graph Neural Network (GNN) regressors on eleven graph datasets. Results show that MP choice significantly impacts performance, but the best layer depends entirely on the specific graph size, training set size, and regression goal. Classical layers often outperform attention-based ones for small graphs at lower costs.