Fig. 10: Behavior and performance of feature selection methods.

a Simulated average number of descriptors selected by each statistical learning method over 100 trials with 100 observations, 1000 candidate descriptors, and a correlation of ρ = 0.5 between the descriptors. b The average number of descriptors selected by each method using 50 randomly separated training set of metal binding energies on dopant-modified or adsorbate-modified MgO surfaces. The error bars represent the range of one standard deviation.