Fig. 5
From: Efficient differentially private learning improves drug sensitivity prediction
Accuracy of drug sensitivity prediction increases with amount of private data for the proposed robust private linear regression. The state-of-the-art methods fail to improve over just using the non-private data under strict privacy conditions, with reasonable data amounts. The baselines (horizontal dashed lines) are learned on 10 non-private data points; the private algorithms additionally have privacy-protected data (x-axis). The non-private algorithm (LR) has the same amount of additional non-privacy-protected data. Accuracy is measured by Spearman’s rank correlation coefficient over ranking cell lines by sensitivity to a drug (left; higher is better) and by weighted probabilistic concordance index (wpc-index; right; higher is better). All methods use 10-dimensional data except the gray baseline showing the best performance with 10 non-private 64-dimensional data points. Private methods use ε=2. Corresponding results for ε=1 are in Additional file 1: Figure S3 and results including non-private robust LR in Additional file 1: Figure S4. The results are averaged over all drugs and 50-fold Monte Carlo cross-validation; error bars denote standard deviation over 50 Monte Carlo repeats