Merge pull request scikit-learn#2829 from arjoly/maxfeatures-seamantics

[MRG] Uniformize max_features semantics for extra trees and random forest

Author: arjoly

doc/whats_new.rst

@@ -20,14 +20,14 @@ Changelog
      :class:`ensemble.BaggingRegressor` meta-estimators for ensembling
      any kind of base estimator. See the :ref:`Bagging <bagging>` section of
      the user guide for details and examples. By `Gilles Louppe`_.
-     
+
    - Memory improvements of decision trees, by `Arnaud Joly`_.
-     
+
    - Decision trees can now be built in best-first manner by using ``max_leaf_nodes``
      as the stopping criteria. Refactored the tree code to use either a
      stack or a priority queue for tree building.
      By `Peter Prettenhofer`_ and `Gilles Louppe`_.
-     
+
    - Decision trees can now be fitted on fortran- and c-style arrays, and
      non-continuous arrays without the need to make a copy.
      If the input array has a different dtype than ``np.float32``, a fortran-
@@ -42,24 +42,24 @@ Changelog
    - Changed the internal storage of decision trees to use a struct array.
      This fixed some small bugs, while improving code and providing a small
      speed gain. By `Joel Nothman`_.
-     
+
    - Reduce memory usage and overhead when fitting and predicting with forests
      of randomized trees in parallel with ``n_jobs != 1`` by leveraging new
      threading backend of joblib 0.8 and releasing the GIL in the tree fitting
      Cython code.  By `Olivier Grisel`_ and `Gilles Louppe`_.
 
    - Speed improvement of the :mod:`sklearn.ensemble.gradient_boosting` module.
      By `Gilles Louppe`_ and `Peter Prettenhofer`_.
-     
+
    - Various enhancements to the  :mod:`sklearn.ensemble.gradient_boosting`
      module: a ``warm_start`` argument to fit additional trees,
      a ``max_leaf_nodes`` argument to fit GBM style trees,
      a ``monitor`` fit argument to inspect the estimator during training, and
      refactoring of the verbose code. By `Peter Prettenhofer`_.
-     
+
    - Fixed bug in :class:`gradient_boosting.GradientBoostingRegressor` with
      ``loss='huber'``: ``gamma`` might have not been initialized.
-     
+
    - Fixed feature importances as computed with a forest of randomized trees
      when fit with ``sample_weight != None`` and/or with ``bootstrap=True``.
      By `Gilles Louppe`_.
@@ -197,6 +197,27 @@ API changes summary
      of alphas was not computed correctly and the scaling with normalize
      was wrong. By `Manoj Kumar`_.
 
+   - Fix wrong maximal number of features drawn (`max_features`) at each split
+     for decision trees, random forests and gradient tree boosting.
+     Previously, the count for the number of drawn features started only after
+     one non constant features in the split. This bug fix will affect
+     computational and generalization performance of those algorithms in the
+     presence of constant features. To get back previous generalization
+     performance, you should modify the value of `max_features`.
+     By `Arnaud Joly`_.
+
+   - Fix wrong maximal number of features drawn (`max_features`) at each split
+     for :class:`ensemble.ExtraTreesClassifier` and
+     :class:`ensemble.ExtraTreesRegressor`. Previously, only non constant
+     features in the split was counted as drawn. Now constant features are
+     counted as drawn. Furthermore at least one feature must be non constant
+     in order to make a valid split. This bug fix will affect
+     computational and generalization performance of extra trees in the
+     presence of constant features. To get back previous generalization
+     performance, you should modify the value of `max_features`.
+     By `Arnaud Joly`_.
+
+
 .. _changes_0_14:
 
 0.14