float numbers can't be set to RFECV's parameter "step"

[lijierr]

Description

When I use RFECV with parameter 'step' as a float number will cause warnings/errors "rfe.py:203: VisibleDeprecationWarning: using a non-integer number instead of an integer will result in an error in the future". And the analysis can't be finished until integer or 1/2.

I read description of RFECV and learned that parameter 'step' can accept float. (introduction online: If greater than or equal to 1, then step corresponds to the (integer) number of features to remove at each iteration. If within (0.0, 1.0), then step corresponds to the percentage (rounded down) of features to remove at each iteration.)

And I didn't read any bugs from source script. Please tell.

[lesteve]

What would be best is if you could provide a standalone snippet reproducing the problem.

[jnothman]

It's actually pretty easy to see in the code where this comes about. The
float semantics are handled for RFE, but not properly for RFECV. This is a
bug. PR welcome.

On 22 September 2016 at 22:16, Loïc Estève notifications@github.com wrote:

What would be best is if you could provide a standalone snippet
reproducing the problem.

—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub
#7467 (comment),
or mute the thread
https://github.com/notifications/unsubscribe-auth/AAEz69klwjPadKdidokVOuzcInf3oHwEks5qsnGxgaJpZM4KDpAQ
.

[naoyak]

Isn't it OK for RFECV.fit() to omit the step validation, since the underlying RFE.fit() call checks this for each fold iteration?

It sounds like the warning is caused due to indexing by float, these lines in 0.17.x might be causing CFE to be initialized with n_features_to_select as float. @joanna-li what version are you on?

[jnothman]

Step is used multiple times. Not all are through RFE

[PedroGFonseca]

Can reproduce on sklearn '0.18.dev0'.

Code sample (works fine with RFE and fails with RFECV):

from sklearn.feature_selection import RFECV

selector = RFECV(model_classifier, step=0.1, cv=3, 
                 scoring='roc_auc', verbose=10, n_jobs=-1)

selector.fit(X_train, y_train)

Symptoms:

• Process goes on forever without ending.
• No cores seem to activate.
• Stdout prints: ... lib/python3.4/site-packages/sklearn/feature_selection/rfe.py:198: DeprecationWarning: using a non-integer number instead of an integer will result in an error in the future support_[features[ranks][:threshold]] = False
• Stdout also prints:
  Fitting estimator with 177 features. Fitting estimator with 177 features. Fitting estimator with 177 features. Fitting estimator with 124 features. Fitting estimator with 124 features. Fitting estimator with 71 features. Fitting estimator with 71 features. Fitting estimator with 124 features. Fitting estimator with 18 features. Fitting estimator with 18 features. Fitting estimator with 71 features. Fitting estimator with 18 features.
  (but nothing actually happens)
• Works fine with integer step

[jnothman]

That's a lot of symptoms! Yes, we know it's not fixed.

[NelleV]

Fixed in #7469