FIX Fallback to ball_tree using minkowski with w for kd_tree (#22241)

jjerphan · ogrisel · lorentzenchr · web-flow · commit 6145cae3d1ce · 2022-01-22T13:52:02.000-05:00
Co-authored-by: Olivier Grisel &lt;olivier.grisel@ensta.org&gt;
Co-authored-by: Christian Lorentzen &lt;lorentzen.ch@gmail.com&gt;
Co-authored-by: Thomas J. Fan &lt;thomasjpfan@gmail.com&gt;
diff --git a/sklearn/neighbors/_base.py b/sklearn/neighbors/_base.py
@@ -535,7 +535,21 @@ def _fit(self, X, y=None):
             ):
                 self._fit_method = "brute"
             else:
-                if self.effective_metric_ in VALID_METRICS["kd_tree"]:
+                if (
+                    self.effective_metric_ == "minkowski"
+                    and self.effective_metric_params_.get("w") is not None
+                ):
+                    # Be consistent with scipy 1.8 conventions: in scipy 1.8,
+                    # 'wminkowski' was removed in favor of passing a
+                    # weight vector directly to 'minkowski'.
+                    #
+                    # 'wminkowski' is not part of valid metrics for KDTree but
+                    # the 'minkowski' without weights is.
+                    #
+                    # Hence, we detect this case and choose BallTree
+                    # which supports 'wminkowski'.
+                    self._fit_method = "ball_tree"
+                elif self.effective_metric_ in VALID_METRICS["kd_tree"]:
                     self._fit_method = "kd_tree"
                 elif (
                     callable(self.effective_metric_)
@@ -553,6 +567,16 @@ def _fit(self, X, y=None):
                 **self.effective_metric_params_,
             )
         elif self._fit_method == "kd_tree":
+            if (
+                self.effective_metric_ == "minkowski"
+                and self.effective_metric_params_.get("w") is not None
+            ):
+                raise ValueError(
+                    "algorithm='kd_tree' is not valid for "
+                    "metric='minkowski' with a weight parameter 'w': "
+                    "try algorithm='ball_tree' "
+                    "or algorithm='brute' instead."
+                )
             self._tree = KDTree(
                 X,
                 self.leaf_size,
diff --git a/sklearn/neighbors/tests/test_neighbors.py b/sklearn/neighbors/tests/test_neighbors.py
@@ -22,7 +22,7 @@
 from sklearn.metrics.pairwise import pairwise_distances
 from sklearn.model_selection import cross_val_score
 from sklearn.model_selection import train_test_split
-from sklearn.neighbors import VALID_METRICS_SPARSE, VALID_METRICS
+from sklearn.neighbors import VALID_METRICS_SPARSE
 from sklearn.neighbors._base import _is_sorted_by_data, _check_precomputed
 from sklearn.pipeline import make_pipeline
 from sklearn.utils._testing import assert_array_almost_equal
@@ -58,6 +58,46 @@
 neighbors.radius_neighbors_graph = ignore_warnings(neighbors.radius_neighbors_graph)
 
 
+def _generate_test_params_for(metric: str, n_features: int):
+    """Return list of dummy DistanceMetric kwargs for tests."""
+
+    # Distinguishing on cases not to compute unneeded datastructures.
+    rng = np.random.RandomState(1)
+    weights = rng.random_sample(n_features)
+
+    if metric == "minkowski":
+        minkowski_kwargs = [dict(p=1.5), dict(p=2), dict(p=3), dict(p=np.inf)]
+        if sp_version >= parse_version("1.8.0.dev0"):
+            # TODO: remove the test once we no longer support scipy < 1.8.0.
+            # Recent scipy versions accept weights in the Minkowski metric directly:
+            # type: ignore
+            minkowski_kwargs.append(dict(p=3, w=rng.rand(n_features)))
+        return minkowski_kwargs
+
+    # TODO: remove this case for "wminkowski" once we no longer support scipy < 1.8.0.
+    if metric == "wminkowski":
+        weights /= weights.sum()
+        wminkowski_kwargs = [dict(p=1.5, w=weights)]
+        if sp_version < parse_version("1.8.0.dev0"):
+            # wminkowski was removed in scipy 1.8.0 but should work for previous
+            # versions.
+            wminkowski_kwargs.append(dict(p=3, w=rng.rand(n_features)))
+        return wminkowski_kwargs
+
+    if metric == "seuclidean":
+        return [dict(V=rng.rand(n_features))]
+
+    if metric == "mahalanobis":
+        A = rng.rand(n_features, n_features)
+        # Make the matrix symmetric positive definite
+        VI = A + A.T + 3 * np.eye(n_features)
+        return [dict(VI=VI)]
+
+    # Case of: "euclidean", "manhattan", "chebyshev", "haversine" or any other metric.
+    # In those cases, no kwargs are needed.
+    return [{}]
+
+
 def _weight_func(dist):
     """Weight function to replace lambda d: d ** -2.
     The lambda function is not valid because:
@@ -1385,58 +1425,48 @@ def custom_metric(x1, x2):
 
 # TODO: Remove filterwarnings in 1.3 when wminkowski is removed
 @pytest.mark.filterwarnings("ignore:WMinkowskiDistance:FutureWarning:sklearn")
-def test_valid_brute_metric_for_auto_algorithm():
-    X = rng.rand(12, 12)
+@pytest.mark.parametrize("metric", neighbors.VALID_METRICS["brute"])
+def test_valid_brute_metric_for_auto_algorithm(metric, n_samples=20, n_features=12):
+    # Any valid metric for algorithm="brute" must be a valid for algorithm="auto".
+    # It's the responsibility of the estimator to select which algorithm is likely
+    # to be the most efficient from the subset of the algorithm compatible with
+    # that metric (and params). Worst case is to fallback to algorithm="brute".
+    X = rng.rand(n_samples, n_features)
     Xcsr = csr_matrix(X)
 
-    # check that there is a metric that is valid for brute
-    # but not ball_tree (so we actually test something)
-    assert "cosine" in VALID_METRICS["brute"]
-    assert "cosine" not in VALID_METRICS["ball_tree"]
+    metric_params_list = _generate_test_params_for(metric, n_features)
+
+    if metric == "precomputed":
+        X_precomputed = rng.random_sample((10, 4))
+        Y_precomputed = rng.random_sample((3, 4))
+        DXX = metrics.pairwise_distances(X_precomputed, metric="euclidean")
+        DYX = metrics.pairwise_distances(
+            Y_precomputed, X_precomputed, metric="euclidean"
+        )
+        nb_p = neighbors.NearestNeighbors(n_neighbors=3, metric="precomputed")
+        nb_p.fit(DXX)
+        nb_p.kneighbors(DYX)
 
-    # Metric which don't required any additional parameter
-    require_params = ["mahalanobis", "wminkowski", "seuclidean"]
-    for metric in VALID_METRICS["brute"]:
-        if metric != "precomputed" and metric not in require_params:
+    else:
+        for metric_params in metric_params_list:
             nn = neighbors.NearestNeighbors(
-                n_neighbors=3, algorithm="auto", metric=metric
-            )
-            if metric != "haversine":
-                nn.fit(X)
-                nn.kneighbors(X)
-            else:
-                nn.fit(X[:, :2])
-                nn.kneighbors(X[:, :2])
-        elif metric == "precomputed":
-            X_precomputed = rng.random_sample((10, 4))
-            Y_precomputed = rng.random_sample((3, 4))
-            DXX = metrics.pairwise_distances(X_precomputed, metric="euclidean")
-            DYX = metrics.pairwise_distances(
-                Y_precomputed, X_precomputed, metric="euclidean"
+                n_neighbors=3,
+                algorithm="auto",
+                metric=metric,
+                metric_params=metric_params,
             )
-            nb_p = neighbors.NearestNeighbors(n_neighbors=3)
-            nb_p.fit(DXX)
-            nb_p.kneighbors(DYX)
+            # Haversine distance only accepts 2D data
+            if metric == "haversine":
+                X = np.ascontiguousarray(X[:, :2])
 
-    for metric in VALID_METRICS_SPARSE["brute"]:
-        if metric != "precomputed" and metric not in require_params:
-            nn = neighbors.NearestNeighbors(
-                n_neighbors=3, algorithm="auto", metric=metric
-            ).fit(Xcsr)
-            nn.kneighbors(Xcsr)
-
-    # Metric with parameter
-    VI = np.dot(X, X.T)
-    list_metrics = [
-        ("seuclidean", dict(V=rng.rand(12))),
-        ("wminkowski", dict(w=rng.rand(12))),
-        ("mahalanobis", dict(VI=VI)),
-    ]
-    for metric, params in list_metrics:
-        nn = neighbors.NearestNeighbors(
-            n_neighbors=3, algorithm="auto", metric=metric, metric_params=params
-        ).fit(X)
-        nn.kneighbors(X)
+            nn.fit(X)
+            nn.kneighbors(X)
+
+            if metric in VALID_METRICS_SPARSE["brute"]:
+                nn = neighbors.NearestNeighbors(
+                    n_neighbors=3, algorithm="auto", metric=metric
+                ).fit(Xcsr)
+                nn.kneighbors(Xcsr)
 
 
 def test_metric_params_interface():
