Feature importance for random forests

[mjmckp]

When training a random forest, the scikit-learn package returns the feature importances (see http://scikit-learn.org/stable/modules/generated/sklearn.ensemble.RandomForestClassifier.html), where the feature importance metric is the mean decrease impurity (as described here http://blog.datadive.net/selecting-good-features-part-iii-random-forests/). Can ML.Net also provide this information?

[codemzs]

Hi @mjmckp , Yes, this information is maintained in the runtime but there isn't a high level API at the moment that exposes feature importance. We are working on exposing model summary via high level API that should contain feature importance.

[zeahmed]

Looking at the new API proposal #371, this feature will be automatically available.

For example, in the proposed example look at this Line where predictor is created. The predictor contains several methods to save models/metadata and SaveSummary is one of them which emits feature importance where applicable.

Also if the predictor implements ICanGetSummaryInKeyValuePairs then feature importance can be obtain as a List of KeyValuePairs. For example, see here for reference.

[shauheen]

Closing this issue as #371 would address the main problem.