I would like to take the top important ones and want to use feature_importances_ for that. The problem is that feature_importances_ is array without reference to the tree nodes. I have the original features but as each one can be more than one time in the tree I'm not sure how to relate importance to node.
I think you are mixing two different things here.
- The feature_importance_ - this is an array which reflects how much each of the model's original features contributes to overall classification quality.
- The features positions in the tree - this is a mere representation of the decision rules made in each step in the tree. A feature position(s) in the tree in terms of importance is not so trivial.
There are some potential heuristics for understanding the relation between the two. If a feature dosn't appear in the tree it has 0 importance and generally the higher the feature is in the tree the more important it is (assuming its being compared to another feature on the same branch).