FIX: Change limits of power_t param to [0, inf) (scikit-learn#31474)

Author: ritvi-alagusankar

doc/whats_new/upcoming_changes/sklearn.linear_model/31474.api.rst

@@ -0,0 +1,6 @@
+- :class:`linear_model.SGDClassifier`, :class:`linear_model.SGDRegressor`, and
+  :class:`linear_model.SGDOneClassSVM` now deprecate negative values for the
+  `power_t` parameter. Using a negative value will raise a warning in version 1.8
+  and will raise an error in version 1.10. A value in the range [0.0, inf) must be used
+  instead.
+  By :user:`Ritvi Alagusankar <ritvi-alagusankar>`

sklearn/linear_model/_stochastic_gradient.py

@@ -731,6 +731,15 @@ def _fit(
                 ),
                 ConvergenceWarning,
             )
+
+        if self.power_t < 0:
+            warnings.warn(
+                "Negative values for `power_t` are deprecated in version 1.8 "
+                "and will raise an error in 1.10. "
+                "Use values in the range [0.0, inf) instead.",
+                FutureWarning,
+            )
+
         return self
 
     def _fit_binary(self, X, y, alpha, C, sample_weight, learning_rate, max_iter):
@@ -1082,7 +1091,11 @@ class SGDClassifier(BaseSGDClassifier):
 
     power_t : float, default=0.5
         The exponent for inverse scaling learning rate.
-        Values must be in the range `(-inf, inf)`.
+        Values must be in the range `[0.0, inf)`.
+
+        .. deprecated:: 1.8
+            Negative values for `power_t` are deprecated in version 1.8 and will raise
+            an error in 1.10. Use values in the range [0.0, inf) instead.
 
     early_stopping : bool, default=False
         Whether to use early stopping to terminate training when validation
@@ -1585,6 +1598,14 @@ def _fit(
                 ConvergenceWarning,
             )
 
+        if self.power_t < 0:
+            warnings.warn(
+                "Negative values for `power_t` are deprecated in version 1.8 "
+                "and will raise an error in 1.10. "
+                "Use values in the range [0.0, inf) instead.",
+                FutureWarning,
+            )
+
         return self
 
     @_fit_context(prefer_skip_nested_validation=True)
@@ -1880,7 +1901,11 @@ class SGDRegressor(BaseSGDRegressor):
 
     power_t : float, default=0.25
         The exponent for inverse scaling learning rate.
-        Values must be in the range `(-inf, inf)`.
+        Values must be in the range `[0.0, inf)`.
+
+        .. deprecated:: 1.8
+            Negative values for `power_t` are deprecated in version 1.8 and will raise
+            an error in 1.10. Use values in the range [0.0, inf) instead.
 
     early_stopping : bool, default=False
         Whether to use early stopping to terminate training when validation
@@ -2118,7 +2143,11 @@ class SGDOneClassSVM(OutlierMixin, BaseSGD):
 
     power_t : float, default=0.5
         The exponent for inverse scaling learning rate.
-        Values must be in the range `(-inf, inf)`.
+        Values must be in the range `[0.0, inf)`.
+
+        .. deprecated:: 1.8
+            Negative values for `power_t` are deprecated in version 1.8 and will raise
+            an error in 1.10. Use values in the range [0.0, inf) instead.
 
     warm_start : bool, default=False
         When set to True, reuse the solution of the previous call to fit as
@@ -2490,6 +2519,14 @@ def _fit(
                 ConvergenceWarning,
             )
 
+        if self.power_t < 0:
+            warnings.warn(
+                "Negative values for `power_t` are deprecated in version 1.8 "
+                "and will raise an error in 1.10. "
+                "Use values in the range [0.0, inf) instead.",
+                FutureWarning,
+            )
+
         return self
 
     @_fit_context(prefer_skip_nested_validation=True)

sklearn/linear_model/tests/test_sgd.py

@@ -1,4 +1,5 @@
 import pickle
+import warnings
 from unittest.mock import Mock
 
 import joblib
@@ -507,6 +508,35 @@ def test_sgd_failing_penalty_validation(Estimator):
         clf.fit(X, Y)
 
 
+# TODO(1.10): remove this test
+@pytest.mark.parametrize(
+    "klass",
+    [
+        SGDClassifier,
+        SparseSGDClassifier,
+        SGDRegressor,
+        SparseSGDRegressor,
+        SGDOneClassSVM,
+        SparseSGDOneClassSVM,
+    ],
+)
+def test_power_t_limits(klass):
+    """Check that a warning is raised when `power_t` is negative."""
+
+    # Check that negative values of `power_t` raise a warning
+    clf = klass(power_t=-1.0)
+    with pytest.warns(
+        FutureWarning, match="Negative values for `power_t` are deprecated"
+    ):
+        clf.fit(X, Y)
+
+    # Check that values of 'power_t in range [0, inf) do not raise a warning
+    with warnings.catch_warnings(record=True) as w:
+        clf = klass(power_t=0.5)
+        clf.fit(X, Y)
+    assert len(w) == 0
+
+
 ###############################################################################
 # Classification Test Case