scikit-learn
diff --git a/‎doc/whats_new/v0.22.rst
Lines changed: 13 additions & 0 deletions b/‎doc/whats_new/v0.22.rst
Lines changed: 13 additions & 0 deletions
diff --git a/‎sklearn/model_selection/_search.py
Lines changed: 5 additions & 0 deletions b/‎sklearn/model_selection/_search.py
Lines changed: 5 additions & 0 deletions
diff --git a/‎sklearn/model_selection/tests/test_search.py
Lines changed: 49 additions & 0 deletions b/‎sklearn/model_selection/tests/test_search.py
Lines changed: 49 additions & 0 deletions
@@ -740,6 +740,19 @@ Changelog
   :class:`preprocessing.KernelCenterer`
   :pr:`14336` by :user:`Gregory Dexter <gdex1>`.
 
+:mod:`sklearn.model_selection`
+..................
+
+- |Fix| :class:`model_selection.GridSearchCV` and
+  `model_selection.RandomizedSearchCV` now supports the
+  :term:`_pairwise` property, which prevents an error during cross-validation
+  for estimators with pairwise inputs (such as
+  :class:`neighbors.KNeighborsClassifier` when :term:`metric` is set to
+  'precomputed').
+  :pr:`13925` by :user:`Isaac S. Robson <isrobson>` and :pr:`15524` by
+  :user:`Xun Tang <xun-tang>`.
+
+
 :mod:`sklearn.svm`
 ..................
 
 
@@ -414,6 +414,11 @@ def __init__(self, estimator, scoring=None, n_jobs=None, iid='deprecated',
     def _estimator_type(self):
         return self.estimator._estimator_type
 
+    @property
+    def _pairwise(self):
+        # allows cross-validation to see 'precomputed' metrics
+        return getattr(self.estimator, '_pairwise', False)
+
     def score(self, X, y=None):
         """Returns the score on the given data, if the estimator has been refit.
 
 
@@ -56,11 +56,13 @@
 from sklearn.tree import DecisionTreeClassifier
 from sklearn.cluster import KMeans
 from sklearn.neighbors import KernelDensity
+from sklearn.neighbors import KNeighborsClassifier
 from sklearn.metrics import f1_score
 from sklearn.metrics import recall_score
 from sklearn.metrics import accuracy_score
 from sklearn.metrics import make_scorer
 from sklearn.metrics import roc_auc_score
+from sklearn.metrics.pairwise import euclidean_distances
 from sklearn.impute import SimpleImputer
 from sklearn.pipeline import Pipeline
 from sklearn.linear_model import Ridge, SGDClassifier, LinearRegression
@@ -1798,3 +1800,50 @@ def get_n_splits(self, *args, **kw):
                              'inconsistent results. Expected \\d+ '
                              'splits, got \\d+'):
         ridge.fit(X[:train_size], y[:train_size])
+
+
+def test_search_cv__pairwise_property_delegated_to_base_estimator():
+    """
+    Test implementation of BaseSearchCV has the _pairwise property
+    which matches the _pairwise property of its estimator.
+    This test make sure _pairwise is delegated to the base estimator.
+
+    Non-regression test for issue #13920.
+    """
+    est = BaseEstimator()
+    attr_message = "BaseSearchCV _pairwise property must match estimator"
+
+    for _pairwise_setting in [True, False]:
+        setattr(est, '_pairwise', _pairwise_setting)
+        cv = GridSearchCV(est, {'n_neighbors': [10]})
+        assert _pairwise_setting == cv._pairwise, attr_message
+
+
+def test_search_cv__pairwise_property_equivalence_of_precomputed():
+    """
+    Test implementation of BaseSearchCV has the _pairwise property
+    which matches the _pairwise property of its estimator.
+    This test ensures the equivalence of 'precomputed'.
+
+    Non-regression test for issue #13920.
+    """
+    n_samples = 50
+    n_splits = 2
+    X, y = make_classification(n_samples=n_samples, random_state=0)
+    grid_params = {'n_neighbors': [10]}
+
+    # defaults to euclidean metric (minkowski p = 2)
+    clf = KNeighborsClassifier()
+    cv = GridSearchCV(clf, grid_params, cv=n_splits)
+    cv.fit(X, y)
+    preds_original = cv.predict(X)
+
+    # precompute euclidean metric to validate _pairwise is working
+    X_precomputed = euclidean_distances(X)
+    clf = KNeighborsClassifier(metric='precomputed')
+    cv = GridSearchCV(clf, grid_params, cv=n_splits)
+    cv.fit(X_precomputed, y)
+    preds_precomputed = cv.predict(X_precomputed)
+
+    attr_message = "GridSearchCV not identical with precomputed metric"
+    assert (preds_original == preds_precomputed).all(), attr_message