BUG in RadiusNeighborClassifier outlier handling

Ruggedized the test, which wasn't doing much previously, and added a test
for c0d4015.

Also replaced flatten (which copies) with ravel and improved the error
message for outliers in the face of outlier_label=None.

Author: larsmans

sklearn/neighbors/classification.py

@@ -289,22 +289,24 @@ def predict(self, X):
         pred_labels = [self._y[ind] for ind in neigh_ind]
 
         if self.outlier_label is not None:
-            outlier_label = np.array((self.outlier_label, ))
-            small_value = np.array((1e-6, ))
+            outlier_label = np.array([self.outlier_label])
+            small_value = np.array([1e-6])
             for i, pl in enumerate(pred_labels):
                 # Check that all have at least 1 neighbor
                 if len(pl) < 1:
                     pred_labels[i] = outlier_label
                     neigh_dist[i] = small_value
         else:
-            for pl in pred_labels:
+            for i, pl in enumerate(pred_labels):
                 # Check that all have at least 1 neighbor
+                # TODO we should gather all outliers, or the first k,
+                # before constructing the error message.
                 if len(pl) < 1:
-                    raise ValueError('no neighbors found for a test sample, '
+                    raise ValueError('No neighbors found for test sample %d, '
                                      'you can try using larger radius, '
                                      'give a label for outliers, '
-                                     'or consider removing them in your '
-                                     'dataset')
+                                     'or consider removing it from your '
+                                     'dataset.' % i)
 
         weights = _get_weights(neigh_dist, self.weights)
 
@@ -316,10 +318,10 @@ def predict(self, X):
                              for (pl, w) in zip(pred_labels, weights)],
                             dtype=np.int)
 
-        mode = mode.flatten().astype(np.int)
+        mode = mode.ravel().astype(np.int)
         # map indices to classes
         prediction = self.classes_.take(mode)
         if self.outlier_label is not None:
             # reset outlier label
-            prediction[mode == outlier_label] = self.outlier_label
+            prediction[prediction == outlier_label] = self.outlier_label
         return prediction

sklearn/neighbors/tests/test_neighbors.py

@@ -240,14 +240,20 @@ def test_radius_neighbors_classifier_when_no_neighbors():
 
     weight_func = _weight_func
 
-    for algorithm in ALGORITHMS:
-        for weights in ['uniform', 'distance', weight_func]:
-            clf = neighbors.RadiusNeighborsClassifier(radius=radius,
-                                                      weights=weights,
-                                                      algorithm=algorithm)
-            clf.fit(X, y)
-            clf.predict(z1)
-            assert_raises(ValueError, clf.predict, z2)
+    for outlier_label in [0, -1, None]:
+        for algorithm in ALGORITHMS:
+            for weights in ['uniform', 'distance', weight_func]:
+                rnc = neighbors.RadiusNeighborsClassifier
+                clf = rnc(radius=radius, weights=weights, algorithm=algorithm,
+                          outlier_label=outlier_label)
+                clf.fit(X, y)
+                assert_array_equal(np.array([1, 2]),
+                                   clf.predict(z1))
+                if outlier_label is None:
+                    assert_raises(ValueError, clf.predict, z2)
+                elif False:
+                    assert_array_equal(np.array([1, outlier_label]),
+                                       clf.predict(z2))
 
 
 def test_radius_neighbors_classifier_outlier_labeling():