Fig. 3 |. Explaining ensembles helps overcome instability in feature discovery performance for single models.

a,b, Relationship between test error and feature discovery performance on bootstrap resampled versions of two synthetic datasets. Both datasets had clusters of highly correlated features; one had a step function outcome (a) and the second had a multiplicative outcome (b). Although there is a high overall correlation between test error and feature discovery performance for both datasets, there is no significant correlation after conditioning on model class (see Supplementary Table 1 for full statistical comparisons across model classes including XGBoost models (GBMs), multilayer perceptron neural network models (MLPs) and elastic net regression). c, Comparison of feature discovery performance between individual models and ensemble models using synthetic and semi-synthetic datasets from our benchmark. The boxes mark the quartiles (25th, 50th and 75th percentiles) of the distribution, and the whiskers extend to show the minimum and maximum of the distribution (excluding outliers). Results for the rest of the datasets and for additional feature attribution methods can be found in Extended Data Figs. 5 and 6.