FIX Changed VarianceThreshold behaviour when threshold is zero. See #13691 (#13704)

rlms · jnothman · commit be03467b9cdc · 2019-05-29T01:54:23.000+10:00
diff --git a/doc/whats_new/v0.22.rst b/doc/whats_new/v0.22.rst
@@ -65,16 +65,21 @@ Changelog
   ``decision_function_shape='ovr'``, and the number of target classes > 2.
   :pr:`12557` by `Adrin Jalali`_.
 
-
 :mod:`sklearn.cluster`
 ..................
 
 - |Enhancement| :class:`cluster.SpectralClustering` now accepts a ``n_components`` 
   parameter. This parameter extends `SpectralClustering` class functionality to
   match `spectral_clustering`.
   :pr:`13726` by :user:`Shuzhe Xiao <fdas3213>`.
+
+:mod:`sklearn.feature_selection`
+................................
+- |Fix| Fixed a bug where :class:`VarianceThreshold` with `threshold=0` did not
+  remove constant features due to numerical instability, by using range
+  rather than variance in this case.
+  :pr:`13704` by `Roddy MacSween <rlms>`.
   
-	
 Miscellaneous
 .............
 
diff --git a/sklearn/feature_selection/tests/test_variance_threshold.py b/sklearn/feature_selection/tests/test_variance_threshold.py
@@ -1,3 +1,6 @@
+import numpy as np
+import pytest
+
 from sklearn.utils.testing import (assert_array_equal, assert_equal,
                                    assert_raises)
 
@@ -26,3 +29,17 @@ def test_variance_threshold():
     for X in [data, csr_matrix(data)]:
         X = VarianceThreshold(threshold=.4).fit_transform(X)
         assert_equal((len(data), 1), X.shape)
+
+
+def test_zero_variance_floating_point_error():
+    # Test that VarianceThreshold(0.0).fit eliminates features that have
+    # the same value in every sample, even when floating point errors
+    # cause np.var not to be 0 for the feature.
+    # See #13691
+
+    data = [[-0.13725701]] * 10
+    assert np.var(data) != 0
+    for X in [data, csr_matrix(data), csc_matrix(data), bsr_matrix(data)]:
+        msg = "No feature in X meets the variance threshold 0.00000"
+        with pytest.raises(ValueError, match=msg):
+            VarianceThreshold().fit(X)
diff --git a/sklearn/feature_selection/variance_threshold.py b/sklearn/feature_selection/variance_threshold.py
@@ -5,7 +5,7 @@
 from ..base import BaseEstimator
 from .base import SelectorMixin
 from ..utils import check_array
-from ..utils.sparsefuncs import mean_variance_axis
+from ..utils.sparsefuncs import mean_variance_axis, min_max_axis
 from ..utils.validation import check_is_fitted
 
 
@@ -65,8 +65,18 @@ def fit(self, X, y=None):
 
         if hasattr(X, "toarray"):   # sparse matrix
             _, self.variances_ = mean_variance_axis(X, axis=0)
+            if self.threshold == 0:
+                mins, maxes = min_max_axis(X, axis=0)
+                peak_to_peaks = maxes - mins
         else:
             self.variances_ = np.var(X, axis=0)
+            if self.threshold == 0:
+                peak_to_peaks = np.ptp(X, axis=0)
+
+        if self.threshold == 0:
+            # Use peak-to-peak to avoid numeric precision issues
+            # for constant features
+            self.variances_ = np.minimum(self.variances_, peak_to_peaks)
 
         if np.all(self.variances_ <= self.threshold):
             msg = "No feature in X meets the variance threshold {0:.5f}"
