show_weights with OneVsRestClassifier about eli5

[kmike]

Could you try passing clf directly, not clf.estimator? eli5 should support OneVsRestClassifier directly; it is a bug if it doesn't work.

By the way, is there a reason you're using OneVsRestClassifier with LogisticRegression? If I'm not mistaken, LR uses OvR internally by default.

[s-j]

Wow, thanks for a quick reply!

I am new to scikit-learn, but to my understanding OneVsRestClassifier adds multi-label support in comparison to pure multi-class. Have I misunderstood this?

Passing clf directly returns "Error: estimator OneVsRestClassifier(estimator=LogisticRegressionCV...) is not supported". I see that there is a dispatch trick in show_prediction, by passing clf as clf.estimator.__class__.

This actually did work for me:

setattr(clf.estimator, 'classes_', clf.classes_)
setattr(clf.estimator, 'coef_', clf.coef_)
setattr(clf.estimator, 'intercept_', clf.intercept_)
eli5.show_weights(clf.estimator, vec=vec, target_names=mlb.classes_)

[kmike]

Ah, right, it looks like OvR is indeed only supported for explain_prediction, not for explain_weights. This looks like an oversight, thanks for the heads up! It must be possible to implement a similar fix for explain_weights. Do you want to send a PR? Alternatively, I can fix that tomorrow.

Yeah, a good point about OneVsRestClassifier and multi-label support; I missed that.

[s-j]

Great, thanks!

[s-j]

(I would appreciate if you could do it. No rush.)

[kmike]

Your example should be fixed by #108. Thanks for the report!