I have built a random forest binary classification model with 18 input variables. Gini importance (MDI) shows that all variables have positive importance, although some are very small. According to permutation importance, all but 3 variables have zero importance. Does that result make sense?
When I set the maximum depth of the forest to 3 or 4, more variables become important for permutation importance.
What should I do to get the "real" importances of the variables?