ENH Use scipy.special.inv_boxcox in PowerTransformer (scikit-learn#27875

)

doc/whats_new/v1.6.rst

@@ -298,6 +298,10 @@ Changelog
   will show the function name in the label.
   :pr:`29158` by :user:`Yao Xiao <Charlie-XIAO>`.
 
+- |Fix| :class:`preprocessing.PowerTransformer` now uses `scipy.special.inv_boxcox`
+  to output `nan` if the input of BoxCox's inverse is invalid.
+  :pr:`27875` by :user:`Xuefeng Xu <xuefeng-xu>`.
+
 :mod:`sklearn.semi_supervised`
 ..............................
 

sklearn/preprocessing/_data.py

@@ -7,7 +7,7 @@
 
 import numpy as np
 from scipy import optimize, sparse, stats
-from scipy.special import boxcox
+from scipy.special import boxcox, inv_boxcox
 
 from ..base import (
     BaseEstimator,
@@ -3376,7 +3376,7 @@ def inverse_transform(self, X):
             X = self._scaler.inverse_transform(X)
 
         inv_fun = {
-            "box-cox": self._box_cox_inverse_tranform,
+            "box-cox": inv_boxcox,
             "yeo-johnson": self._yeo_johnson_inverse_transform,
         }[self.method]
         for i, lmbda in enumerate(self.lambdas_):
@@ -3385,17 +3385,6 @@ def inverse_transform(self, X):
 
         return X
 
-    def _box_cox_inverse_tranform(self, x, lmbda):
-        """Return inverse-transformed input x following Box-Cox inverse
-        transform with parameter lambda.
-        """
-        if lmbda == 0:
-            x_inv = np.exp(x)
-        else:
-            x_inv = (x * lmbda + 1) ** (1 / lmbda)
-
-        return x_inv
-
     def _yeo_johnson_inverse_transform(self, x, lmbda):
         """Return inverse-transformed input x following Yeo-Johnson inverse
         transform with parameter lambda.

sklearn/preprocessing/tests/test_data.py

@@ -2329,6 +2329,11 @@ def test_optimization_power_transformer(method, lmbda):
     n_samples = 20000
     X = rng.normal(loc=0, scale=1, size=(n_samples, 1))
 
+    if method == "box-cox":
+        # For box-cox, means that lmbda * y + 1 > 0 or y > - 1 / lmbda
+        # Clip the data here to make sure the inequality is valid.
+        X = np.clip(X, -1 / lmbda + 1e-5, None)
+
     pt = PowerTransformer(method=method, standardize=False)
     pt.lambdas_ = [lmbda]
     X_inv = pt.inverse_transform(X)
@@ -2341,6 +2346,14 @@ def test_optimization_power_transformer(method, lmbda):
     assert_almost_equal(1, X_inv_trans.std(), decimal=1)
 
 
+def test_invserse_box_cox():
+    # output nan if the input is invalid
+    pt = PowerTransformer(method="box-cox", standardize=False)
+    pt.lambdas_ = [0.5]
+    X_inv = pt.inverse_transform([[-2.1]])
+    assert np.isnan(X_inv)
+
+
 def test_yeo_johnson_darwin_example():
     # test from original paper "A new family of power transformations to
     # improve normality or symmetry" by Yeo and Johnson.