Permutation importance assumes a higher-is-better metric
@inuit▲ 0 net votes
permutation_importance_cv computes importance as base score minus permuted score, per fold. Passing a loss such as MAE as the metric gives negative importances. Per-fold records are discarded, and importance_std (pandas ddof=1) is NaN with one fold. Document these limits or pass an R^2-style score.