The Split Matters: Flat Minima Methods for Improving the Performance of GNNs

Benchmark Model Rank Results
node-classification-on-citeseerGraph-MLP + SWA#4Accuracy: 77.99 ± 1.57%
node-classification-on-citeseer-with-publicGraph-MLP + PGN#4Accuracy: 74.73 ± 0.6%
node-classification-on-coraGAT + SWA#4Accuracy: 88.66 ± 1.38%
node-classification-on-cora-with-public-splitGAT+PGN#19Accuracy: 83.26 ± 0.69%
node-classification-on-ppiGCN + SAF#8F1: 99.38 ± 0.01%
node-classification-on-ppiGAT + PGN#10F1: 99.34 ± 0.02%
node-classification-on-pubmedGraph-MLP + SAF#3Accuracy: 90.64 ± 0.46%
node-classification-on-pubmed-60-20-20-randomGraph-MLP + SAF#81:1 Accuracy: 90.64 ± 0.46%
node-classification-on-pubmed-with-publicGraph-MLP + ASAM#4Accuracy: 82.60 ± 0.80%