Global Self-Attention as a Replacement for Graph Convolution

Benchmark Model Rank Results
graph-classification-on-cifar10-100kEGT#14Accuracy (%): 68.702
graph-classification-on-mnistEGT#9Accuracy: 98.173
graph-property-prediction-on-ogbg-molhivEGT#7Test ROC-AUC: 0.806 ± 0.0065
graph-property-prediction-on-ogbg-molpcbaEGT#11Test AP: 0.2961 ± 0.0024
graph-regression-on-pcqm4m-lscEGT#10Validation MAE: 0.1224
graph-regression-on-pcqm4mv2-lscEGT + Triangular Attention#2Validation MAE: 0.0671Test MAE: 0.0683
graph-regression-on-pcqm4mv2-lscEGT#10Validation MAE: 0.0857Test MAE: 0.0862
graph-regression-on-zinc-100kEGT#4MAE: 0.143
graph-regression-on-zinc-500kEGT#23MAE: 0.108
link-prediction-on-tsp-hcp-benchmark-setEGT#3F1: 0.853
node-classification-on-clusterEGT#2Accuracy: 79.232
node-classification-on-patternEGT#5Accuracy: 86.821
node-classification-on-pattern-100kEGT#1Accuracy (%): 86.816