Skip to main content

Table 3 Differences in best hyperparameter selection for different folding methods on the public data set. The top 10 performing hyperparameter sets for the random fold splitting are given together with the respective rank of this setting for the other folding methods

From: Splitting chemical structure data sets for federated privacy-preserving machine learning

Hidden Dropout Weight Rank Rank Rank
Sizes   Decay Sphere exclusion Scaffold network LSH
[2000] 0.7 1E−6 2 1 8
[2000] 0.6 1E−6 8 6 5
[1600] 0.7 1E−6 3 2 2
[1200] 0.5 1E−6 11 13 7
[1200] 0.6 1E−6 9 10 4
[3000] 0.7 1E−6 5 4 6
[1200] 0.7 1E−6 1 3 1
[1600] 0.6 1E−6 6 9 3
[3000] 0.6 1E−6 7 8 11
[1600] 0.5 1E-6 4 11 10