Fig. 2.

Comparison between SML, iMLE from SML or from majority voting, the best inferred predictor, and the median balanced accuracy of all ensemble predictors on real-world datasets. The ACS dataset (Left) approximately satisfied our assumptions. In the NASDAQ dataset (Center) many predictors had poor performances. For the LASTFM data (Right), the predictors did not satisfy the conditional independence assumption. In all cases iMLE starting from SML had equal or higher balanced accuracy than iMLE starting from majority voting. The boxplots represent the distribution of balanced accuracies over 30 independent runs.