[MRG] Correcting length of explained_variance_ratio_, eigen solver (s…

…cikit-learn#7632)

doc/whats_new.rst

@@ -98,11 +98,31 @@ Bug fixes
  `#6497 <https://github.com/scikit-learn/scikit-learn/pull/6497>`_
  by `Sebastian Säger`_
 
+ - Attribute ``explained_variance_ratio`` of
+ :class:`discriminant_analysis.LinearDiscriminantAnalysis` calculated
+ with SVD and Eigen solver are now of the same length. (`#7632
+ <https://github.com/scikit-learn/scikit-learn/pull/7632>`_).
+ By `JPFrancoia`_
+
  - Fixes issue in :ref:`univariate_feature_selection` where score 
  functions were not accepting multi-label targets.(`#7676
  <https://github.com/scikit-learn/scikit-learn/pull/7676>`_)
  by `Mohammed Affan`_
-
+
+
+API changes summary
+-------------------
+
+Linear, kernelized and related models
+
+ - Length of `explained_variance_ratio` of
+ :class:`discriminant_analysis.LinearDiscriminantAnalysis`
+ changed for both Eigen and SVD solvers. The attribute has now a length
+ of min(n_components, n_classes - 1). (`#7632
+ <https://github.com/scikit-learn/scikit-learn/pull/7632>`_).
+ By `JPFrancoia`_
+
+
 .. _changes_0_18:
 
 Version 0.18
@@ -571,8 +591,8 @@ Decomposition, manifold learning and clustering
  :class:`decomposition.KernelPCA`, :class:`manifold.LocallyLinearEmbedding`,
  and :class:`manifold.SpectralEmbedding` (`#5012 <https://github.com/scikit-learn/scikit-learn/pull/5012>`_). By `Peter Fischer`_.
 
- - Attribute ``explained_variance_ratio_`` calculated with the SVD solver of
- :class:`discriminant_analysis.LinearDiscriminantAnalysis` now returns
+ - Attribute ``explained_variance_ratio_`` calculated with the SVD solver
+ of :class:`discriminant_analysis.LinearDiscriminantAnalysis` now returns
  correct results. By `JPFrancoia`_
 
 Preprocessing and feature selection

sklearn/discriminant_analysis.py

@@ -293,8 +293,8 @@ def _solve_lsqr(self, X, y, shrinkage):
  self.means_ = _class_means(X, y)
  self.covariance_ = _class_cov(X, y, self.priors_, shrinkage)
  self.coef_ = linalg.lstsq(self.covariance_, self.means_.T)[0].T
- self.intercept_ = (-0.5 * np.diag(np.dot(self.means_, self.coef_.T))
- + np.log(self.priors_))
+ self.intercept_ = (-0.5 * np.diag(np.dot(self.means_, self.coef_.T)) +
+ np.log(self.priors_))
 
  def _solve_eigen(self, X, y, shrinkage):
  """Eigenvalue solver.
@@ -336,15 +336,16 @@ class scatter). This solver supports both classification and
  Sb = St - Sw # between scatter
 
  evals, evecs = linalg.eigh(Sb, Sw)
- self.explained_variance_ratio_ = np.sort(evals / np.sum(evals))[::-1]
+ self.explained_variance_ratio_ = np.sort(evals / np.sum(evals)
+ )[::-1][:self._max_components]
  evecs = evecs[:, np.argsort(evals)[::-1]] # sort eigenvectors
  # evecs /= np.linalg.norm(evecs, axis=0) # doesn't work with numpy 1.6
  evecs /= np.apply_along_axis(np.linalg.norm, 0, evecs)
 
  self.scalings_ = evecs
  self.coef_ = np.dot(self.means_, evecs).dot(evecs.T)
- self.intercept_ = (-0.5 * np.diag(np.dot(self.means_, self.coef_.T))
- + np.log(self.priors_))
+ self.intercept_ = (-0.5 * np.diag(np.dot(self.means_, self.coef_.T)) +
+ np.log(self.priors_))
 
  def _solve_svd(self, X, y):
  """SVD solver.
@@ -400,12 +401,12 @@ def _solve_svd(self, X, y):
  _, S, V = linalg.svd(X, full_matrices=0)
 
  self.explained_variance_ratio_ = (S**2 / np.sum(
-  S**2))[:self.n_components]
+ S**2))[:self._max_components]
  rank = np.sum(S > self.tol * S[0])
  self.scalings_ = np.dot(scalings, V.T[:, :rank])
  coef = np.dot(self.means_ - self.xbar_, self.scalings_)
- self.intercept_ = (-0.5 * np.sum(coef ** 2, axis=1)
- + np.log(self.priors_))
+ self.intercept_ = (-0.5 * np.sum(coef ** 2, axis=1) +
+ np.log(self.priors_))
  self.coef_ = np.dot(coef, self.scalings_.T)
  self.intercept_ -= np.dot(self.xbar_, self.coef_.T)
 
@@ -457,6 +458,13 @@ def fit(self, X, y, store_covariance=None, tol=None):
  UserWarning)
  self.priors_ = self.priors_ / self.priors_.sum()
 
+ # Get the maximum number of components
+ if self.n_components is None:
+ self._max_components = len(self.classes_) - 1
+ else:
+ self._max_components = min(len(self.classes_) - 1,
+ self.n_components)
+
  if self.solver == 'svd':
  if self.shrinkage is not None:
  raise NotImplementedError('shrinkage not supported')
@@ -497,9 +505,8 @@ def transform(self, X):
  X_new = np.dot(X - self.xbar_, self.scalings_)
  elif self.solver == 'eigen':
  X_new = np.dot(X, self.scalings_)
- n_components = X.shape[1] if self.n_components is None \
- else self.n_components
- return X_new[:, :n_components]
+
+ return X_new[:, :self._max_components]
 
  def predict_proba(self, X):
  """Estimate probability.

sklearn/tests/test_discriminant_analysis.py

@@ -171,19 +171,17 @@ def test_lda_explained_variance_ratio():
  clf_lda_eigen = LinearDiscriminantAnalysis(solver="eigen")
  clf_lda_eigen.fit(X, y)
  assert_almost_equal(clf_lda_eigen.explained_variance_ratio_.sum(), 1.0, 3)
+ assert_equal(clf_lda_eigen.explained_variance_ratio_.shape, (2,),
+ "Unexpected length for explained_variance_ratio_")
 
  clf_lda_svd = LinearDiscriminantAnalysis(solver="svd")
  clf_lda_svd.fit(X, y)
  assert_almost_equal(clf_lda_svd.explained_variance_ratio_.sum(), 1.0, 3)
+ assert_equal(clf_lda_svd.explained_variance_ratio_.shape, (2,),
+ "Unexpected length for explained_variance_ratio_")
 
- tested_length = min(clf_lda_svd.explained_variance_ratio_.shape[0],
- clf_lda_eigen.explained_variance_ratio_.shape[0])
-
- # NOTE: clf_lda_eigen.explained_variance_ratio_ is not of n_components
- # length. Make it the same length as clf_lda_svd.explained_variance_ratio_
- # before comparison.
  assert_array_almost_equal(clf_lda_svd.explained_variance_ratio_,
- clf_lda_eigen.explained_variance_ratio_[:tested_length])
+ clf_lda_eigen.explained_variance_ratio_)
 
 
 def test_lda_orthogonality():