Cross‐validated permutation feature importance considering correlation between features

Abstract In molecular design, material process and control, it is important not only to construct a model with high predictive ability between explanatory features x objective y using dataset but also interpret the constructed model. An index of feature importance in permutation (PFI), which can be combined any regressors classifiers. However, PFI becomes unstable when number samples low because necessary divide into training validation data calculating it. Additionally, there are strongly correlated x, these estimated low. Hence, cross‐validated (CVPFI) method proposed. CVPFI calculated stably, even small samples, construction evaluation repeated based on cross‐validation. Furthermore, by considering absolute correlation coefficients features, evaluated appropriately x. Case studies numerical simulation actual compound showed that compared PFI. This possible low, linear nonlinear relationships mixed strong correlations quantised biased exist Python codes for available at https://github.com/hkaneko1985/dcekit .