Fig. 4

Correlation between predictive power and dataset-specific features for random forest models with dataset-pipeline interactions (A) and pipeline features only (B) and penalized linear regression models with interactions (C) and pipeline features only (D). In each panel, the predictive power is quantified by the correlation between model predictions and the ARI of the clusters and the ground truth (top, ARI) or by the correlation between model predictions and the true metric values (bottom, metric). Only features that are significantly associated with predictive power for at least one metric are shown