Accuracy changes with missing features. Top: comparing the drop in accuracy from a typical (30% split) validation set without missing data to one where missing data have been generated by randomly selecting two features per object and replacing them with the corresponding mean values (purple circles) or the values imputed by the iterative imputer (green pentagons). The mean value obtained for the imputer is less than 0.1 and almost three times better than the mean drop for mean values. Bottom: iterative imputer, which is more capable of handling an increased number of missing features, with a limit at three (out of five available in total). The loss in accuracy is less than 20%.