multiclass oversampling

imblearn/over_sampling/_rose.py

@@ -4,12 +4,13 @@
 from scipy import sparse
 
 from sklearn.utils import check_random_state
+from sklearn.utils.multiclass import type_of_target
 
 from .base import BaseOverSampler
 
-from ..utils import check_target_type
+#from ..utils import check_target_type
 from ..utils._validation import _deprecate_positional_args
-from sklearn.utils.multiclass import type_of_target
+
 
 class ROSE(BaseOverSampler):
 
@@ -58,7 +59,9 @@ def _make_samples(self,
  n_class_samples,
  h_shrink):
  """ A support function that returns artificial samples constructed from
- FIXME
+ a random subsample of the data, by adding a multiviariate gaussian kernel
+ and sampling from this distribution. An optional shrink factor can be
+ included, to compress/dilate the kernel.
 
  Parameters
  ----------
@@ -83,18 +86,22 @@ def _make_samples(self,
  Target values for synthetic samples.
 
  """
- p = X.shape[1]
-
+ # get number of features
+ number_of_features = X.shape[1]
+ # import random state from API
  random_state = check_random_state(self.random_state)
- samples_indices = random_state.choice(
- class_indices, size=n_class_samples, replace=True)
-
- h_opt = (4 / ((p + 2) * len(class_indices))) ** (1 / (p + 4))
- H_opt = h_shrink * h_opt * np.diagflat(
- X[class_indices, :].std(axis=0, ddof=1))
-
+ # get random subsample of data with replacement
+ samples_indices = random_state.choice(class_indices, size=n_class_samples, replace=True)
+ # compute optimal min(AMISE)
+ minimize_amise = (4 / ((number_of_features + 2) * len(
+ class_indices))) ** (1 / (number_of_features + 4))
+ # create a diagonal matrix with the st.dev. of all classes
+ variances = np.diagflat(X[class_indices, :].std(axis=0, ddof=1))
+ # compute H_opt = optional shrink factors * min(AMISE) * kernel variance
+ h_opt = h_shrink * minimize_amise * variances
+ # (sample from multivariate normal)* h_opt + original values
  Xrose = np.random.standard_normal(
- size=(n_class_samples, p)) @ H_opt + X[samples_indices, :]
+ size=(n_class_samples, number_of_features)) @ h_opt + X[samples_indices, :]
 
  return Xrose
 
@@ -112,8 +119,11 @@ def _fit_resample(self, X, y):
  }
 
  for class_sample, n_samples in self.sampling_strategy_.items():
+ # get indices of all y's with a given class n
  class_indices = np.flatnonzero(y == class_sample)
+ # compute final n. of samples, by number of elements of a class + n_samples
  n_class_samples = len(class_indices) + n_samples
+ # resample
  X_new = self._make_samples(X,
  class_indices,
  n_class_samples,