-
Notifications
You must be signed in to change notification settings - Fork 34
Open
Description
Based on a CMSTalk conversation there are several methods for feature importance we may wish to add.
Methods:
- Taylor expansion (usually up to second order) of the NN output function w.r.t. the input features. Can be done during (learn when a feature becomes important) or after training (learn the most important features). Already discussed during a journal club. This presentation contains a Jupyter example. The method is described in this paper.
Tools section:
- Several methods are encapsulated in the innvestigate package. Andrzej Novak recommends the "integrated gradients" method.
- The shap module can be used to obtain a ranking of the input features. The method is based on the Shapley values, which originally come from game theory and are computed by integrating-out sets of input features. This is somewhat similar to the "recursive feature elimination" which is already mentioned in the documentation.
Reactions are currently unavailable
Metadata
Metadata
Assignees
Labels
No labels