Is there a way in scikit-learn to compute (or to obtain) the features importance of each class to predict? I known feature_importances_ properties of tree models, but I'm interested in a separate list of more importante predictors for the first class, a separate list for the second class and so on.
E.G: which are the most important features for the Male class? And which are the most important features for the Female class?
The implementation in scikit-learn is based in the formal definition of feature importances given by Friedman [1] in 2001. Since the concept of "feature importance" its somehow fuzzy, Friedman linked his definition to one specific classification method, gradient boosting trees, which by default can handle multi-class classification problems. As far as I know, there is not a formal definition of class-specific importance for any classification method.
That said, I believe you could apply a simple trick to obtain what could be regarded as class-specific feature importances. For instance, imagine you have a classification problem with classes $A$,$B$ and $C$. Instead of training a single gradient boosting trees model for all classes, you could use a one-vs-all [2] strategy and train one model with the samples from $A$ as positive and the samples from $B$,$C$ as negative. Then you train a model with the samples from $B$ as positive and the samples from $A$,$C$ as negative. And finally a classifier with the samples from $C$ as positive and the samples form $A$,$B$ as negative.
After training your models, you could interpret the feature importances reported by the first model as the importances to separate $A$ form the other classes, the importances reported by the second model as the importances to separate $B$ from the rest, and so on.
If your problem has only two classes, then the above approach will yield the same importances vector for both classes, so I won't be very useful. In that case perhaps you might find it useful to compute some representative or prototypical samples from each class (for instance, by using k-Means).
Hope that was of help.
[1] Friedman, J. H. (2001). Greedy function approximation: a gradient boosting machine. Annals of statistics, 1189-1232.
[2] https://en.wikipedia.org/wiki/Multiclass_classification#One-vs.-rest
