DropGNN: Random Dropouts Increase the Expressiveness of Graph Neural Networks

Benchmark Model Rank Results
graph-classification-on-ddDropGIN#30Accuracy: 78.151±3.711
graph-classification-on-enzymesDropGIN#21Accuracy: 65.128±4.117
graph-classification-on-imdb-bDropGIN#18Accuracy: 75.7%
graph-classification-on-imdb-mDropGIN#14Accuracy: 51.4%
graph-classification-on-mutagDropGIN#17Accuracy: 90.4%
graph-classification-on-nci1DropGIN#18Accuracy: 84.331±1.564
graph-classification-on-nci109DropGIN#6Accuracy: 83.961±1.141
graph-classification-on-proteinsDropGIN#42Accuracy: 76.3%
graph-classification-on-ptcDropGIN#15Accuracy: 66.3%
graph-regression-on-esr2DropGIN#3R2: 0.675±0.000RMSE: 0.503±0.675
graph-regression-on-f2DropGIN#5R2: 0.886±0.000RMSE: 0.343±0.886
graph-regression-on-kitGINDrop#3R2: 0.835±0.000RMSE: 0.441±0.835
graph-regression-on-lipophilicityDropGIN#5RMSE: 0.552±0.012R2: 0.809±0.008
graph-regression-on-parp1DropGIN#5R2: 0.920±0.000RMSE: 0.354±0.920
graph-regression-on-pgrGINDrop#3R2: 0.702±0.000RMSE: 0.527±0.702
molecular-property-prediction-on-esolDropGIN#5RMSE: 0.520±0.048R2: 0.935±0.012
molecular-property-prediction-on-freesolvDropGIN#2RMSE: 0.657±0.059R2: 0.972±0.005