Fix PowerTransformer leaves constant feature unchanged (scikit-learn#…

…26566)

Co-authored-by: Olivier Grisel <[email protected]>

doc/whats_new/v1.3.rst

@@ -640,6 +640,11 @@ Changelog
  using `method="box-cox"` on data with a constant `np.nan` column.
  :pr:`26400` by :user:`Yao Xiao <Charlie-XIAO>`.
 
+- |Fix| :class:`preprocessing.PowerTransformer` with `method="yeo-johnson"` now leaves
+ constant features unchanged instead of transforming with an arbitrary value for
+ the `lambdas_` fitted parameter.
+ :pr:`26566` by :user:`Jérémie du Boisberranger <jeremiedbb>`.
+
 :mod:`sklearn.svm`
 ..................
 

sklearn/preprocessing/_data.py

@@ -3150,21 +3150,34 @@ def _fit(self, X, y=None, force_transform=False):
  if not self.copy and not force_transform: # if call from fit()
  X = X.copy() # force copy so that fit does not change X inplace
 
+ n_samples = X.shape[0]
+ mean = np.mean(X, axis=0, dtype=np.float64)
+ var = np.var(X, axis=0, dtype=np.float64)
+
  optim_function = {
  "box-cox": self._box_cox_optimize,
  "yeo-johnson": self._yeo_johnson_optimize,
  }[self.method]
+
+ transform_function = {
+ "box-cox": boxcox,
+ "yeo-johnson": self._yeo_johnson_transform,
+ }[self.method]
+
  with np.errstate(invalid="ignore"): # hide NaN warnings
- self.lambdas_ = np.array([optim_function(col) for col in X.T])
-
- if self.standardize or force_transform:
- transform_function = {
- "box-cox": boxcox,
- "yeo-johnson": self._yeo_johnson_transform,
- }[self.method]
- for i, lmbda in enumerate(self.lambdas_):
- with np.errstate(invalid="ignore"): # hide NaN warnings
- X[:, i] = transform_function(X[:, i], lmbda)
+ self.lambdas_ = np.empty(X.shape[1], dtype=X.dtype)
+ for i, col in enumerate(X.T):
+ # For yeo-johnson, leave constant features unchanged
+ # lambda=1 corresponds to the identity transformation
+ is_constant_feature = _is_constant_feature(var[i], mean[i], n_samples)
+ if self.method == "yeo-johnson" and is_constant_feature:
+ self.lambdas_[i] = 1.0
+ continue
+
+ self.lambdas_[i] = optim_function(col)
+
+ if self.standardize or force_transform:
+ X[:, i] = transform_function(X[:, i], self.lambdas_[i])
 
  if self.standardize:
  self._scaler = StandardScaler(copy=False)

sklearn/preprocessing/tests/test_data.py

@@ -2669,3 +2669,22 @@ def test_kernel_centerer_feature_names_out():
  names_out = centerer.get_feature_names_out()
  samples_out2 = X_pairwise.shape[1]
  assert_array_equal(names_out, [f"kernelcenterer{i}" for i in range(samples_out2)])
+
+
+@pytest.mark.parametrize("standardize", [True, False])
+def test_power_transformer_constant_feature(standardize):
+ """Check that PowerTransfomer leaves constant features unchanged."""
+ X = [[-2, 0, 2], [-2, 0, 2], [-2, 0, 2]]
+
+ pt = PowerTransformer(method="yeo-johnson", standardize=standardize).fit(X)
+
+ assert_allclose(pt.lambdas_, [1, 1, 1])
+
+ Xft = pt.fit_transform(X)
+ Xt = pt.transform(X)
+
+ for Xt_ in [Xft, Xt]:
+ if standardize:
+ assert_allclose(Xt_, np.zeros_like(X))
+ else:
+ assert_allclose(Xt_, X)