minor fix

smote_variants/_version.py

@@ -6,4 +6,4 @@
 @author: gykovacs
 """
 
-__version__= '1.0.0'
+__version__= '1.0.1'

smote_variants/undersampling/_oversampling_driven_undersampling.py

@@ -12,16 +12,18 @@ class OversamplingDrivenUndersampling(UnderSampling):
     The oversampling driven undersampling
     """
 
-    def __init__(self, oversampler_specification, random_state=None):
+    def __init__(self, oversampler_specification, mode="random", random_state=None):
         """
         The constructor of the oversampling driven undersampling
 
         Args:
             oversampler_specification (tuple): the specification of the oversampler
             random_state (None|int|np.random.RandomState): the random seed or state to be used
+            mode (str): 'random'/'farthest' - the mode of sample removal
         """
         UnderSampling.__init__(self, random_state=random_state)
         self.oversampler = instantiate_obj(oversampler_specification)
+        self.mode = mode
 
     @classmethod
     def parameter_combinations(cls, raw=False):
@@ -37,7 +39,8 @@ def parameter_combinations(cls, raw=False):
                 ("smote_variants", "SMOTE", {"random_state": 5}),
                 ("smote_variants", "ADASYN", {"random_state": 5}),
                 ("smote_variants", "Borderline_SMOTE1", {"random_state": 5}),
-            ]
+            ],
+            "mode": ["random", "farthest"],
         }
 
         return cls.generate_parameter_combinations(parameter_combinations, raw)
@@ -63,11 +66,16 @@ def sample(self, X, y):
         dists = X_new - X_maj[:, None]
 
         dists = np.mean(np.sqrt(np.sum((dists) ** 2, axis=2)), axis=1)
-        dists = 1.0 / dists
-        p = dists / np.sum(dists)
 
-        mask = self.random_state.choice(np.arange(n_maj), n_min, p=p, replace=False)
-        X_maj = X_maj[mask]
+        if self.mode == "random":
+            inv_dists = 1.0 / np.where(dists < 1e-8, 1e-8, dists)
+            p = dists / np.sum(inv_dists)
+
+            mask = self.random_state.choice(np.arange(n_maj), n_min, p=p, replace=False)
+            X_maj = X_maj[mask]
+        elif self.mode == "farthest":
+            sorting = np.argsort(dists)
+            X_maj = X_maj[sorting[:n_min]]
 
         X_res = np.vstack([X_maj, X_min])  # pylint: disable=invalid-name
         y_res = np.hstack([np.repeat(0, n_min), np.repeat(1, n_min)])