Fig. 3

Test set (held out dataset) correlations of predicted unsupervised metric values with actual metric values (top) and pipeline-specific adjusted Rand index (ARI) (bottom) for random forest and penalized linear regression models, considering two settings of pipeline features only and dataset-pipeline interactions. In general, including dataset-pipeline interactions improves performance when predicting the unsupervised metrics (top) but degrades performance when predictions are contrasted to pipeline-specific ARI (bottom). P-values are Benjamini–Hochberg multiple test corrected