FIX: bug fix in Nearmiss-3 for not returning the right indices (#282)

Author: glemaitre

doc/whats_new.rst

@@ -9,6 +9,12 @@ Release history
 Changelog
 ---------
 
+Bug fixes
+---------
+
+- Fixed a bug in :class:`under_sampling.NearMiss` version 3. The
+  indices returned were wrong. By `Guillaume Lemaitre`_.
+
 New features
 ~~~~~~~~~~~~
 

imblearn/under_sampling/nearmiss.py

@@ -355,11 +355,11 @@ def _sample(self, X, y):
 
                 # Create the subset containing the samples found during the NN
                 # search. Linearize the indexes and remove the double values
-                idx_vec = np.unique(idx_vec.reshape(-1))
+                idx_vec_farthest = np.unique(idx_vec.reshape(-1))
 
                 # Create the subset
-                sub_samples_x = sub_samples_x[idx_vec, :]
-                sub_samples_y = sub_samples_y[idx_vec]
+                sub_samples_x = sub_samples_x[idx_vec_farthest, :]
+                sub_samples_y = sub_samples_y[idx_vec_farthest]
 
                 # Compute the NN considering the current class
                 dist_vec, idx_vec = self.nn_.kneighbors(
@@ -372,6 +372,10 @@ def _sample(self, X, y):
                     num_samples,
                     key,
                     sel_strategy='farthest')
+
+                # idx_tmp is relative to the feature selected in the
+                # previous step and we need to find the indirection
+                idx_tmp = np.flatnonzero(y == key)[idx_vec_farthest[idx_tmp]]
             else:
                 raise NotImplementedError
 

imblearn/under_sampling/tests/test_nearmiss.py

@@ -10,17 +10,24 @@
 
 # Generate a global dataset to use for the 3 version of nearmiss
 RND_SEED = 0
-X = np.array([[1.17737838, -0.2002118], [0.4960075, 0.86130762],
-              [-0.05903827, 0.10947647], [0.91464286, 1.61369212],
-              [-0.54619583, 1.73009918], [-0.60413357, 0.24628718],
-              [0.45713638, 1.31069295], [-0.04032409, 3.01186964],
-              [0.03142011, 0.12323596], [0.50701028, -0.17636928],
-              [-0.80809175, -1.09917302], [-0.20497017, -0.26630228],
-              [0.99272351, -0.11631728], [-1.95581933, 0.69609604],
+X = np.array([[1.17737838, -0.2002118],
+              [0.4960075, 0.86130762],
+              [-0.05903827, 0.10947647],
+              [0.91464286, 1.61369212],
+              [-0.54619583, 1.73009918],
+              [-0.60413357, 0.24628718],
+              [0.45713638, 1.31069295],
+              [-0.04032409, 3.01186964],
+              [0.03142011, 0.12323596],
+              [0.50701028, -0.17636928],
+              [-0.80809175, -1.09917302],
+              [-0.20497017, -0.26630228],
+              [0.99272351, -0.11631728],
+              [-1.95581933, 0.69609604],
               [1.15157493, -1.2981518]])
 Y = np.array([1, 2, 1, 0, 2, 1, 2, 2, 1, 2, 0, 0, 2, 1, 2])
 
-VERSION_NEARMISS = [1, 2, 3]
+VERSION_NEARMISS = (1, 2, 3)
 
 
 # FIXME remove at the end of the deprecation 0.4
@@ -134,7 +141,7 @@ def test_nm_fit_sample_auto_indices():
 
     idx_gt = [np.array([3, 10, 11, 2, 8, 5, 9, 1, 6]),
               np.array([3, 10, 11, 2, 8, 5, 9, 1, 6]),
-              np.array([3, 10, 11, 0, 2, 3, 5, 1, 4])]
+              np.array([3, 10, 11, 0, 5, 8, 14, 4, 12])]
 
     for version_idx, version in enumerate(VERSION_NEARMISS):
         nm = NearMiss(ratio=ratio, random_state=RND_SEED,