Fig. 3
From: Efficient differentially private learning improves drug sensitivity prediction
Optimal privacy budget split between sufficient statistics. Accuracy on a synthetic data set improves as a bigger proportion of the fixed privacy budget is assigned for \(n\protect \overline {xy}\). The best performance is achieved by assigning term \(n\protect \overline {yy}\) the smallest proportion 5%, term \(n\protect \overline {xy}\) a large 60% proportion, and term \(n\protect \overline {xx}\) the remaining 35% proportion of the privacy budget. Accuracy has been evaluated with 10-dimensional synthetic data, measured by Spearman’s rank correlation between the predicted and true values (higher values are better)