0.8.3

CHANGELOG.md

@@ -1,6 +1,12 @@
 # Changelog of Scikit-Criteria
 
 <!-- BODY -->
+
+## Version 0.8.3
+
+- Fixed a bug detected on the EntropyWeighted, Now works as the literature
+  specifies
+
 ## Version 0.8.2
 
 - We bring back Python 3.7 because is the version used in google.colab.

pyproject.toml

@@ -21,7 +21,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "scikit-criteria"
-version = "0.8.2"
+version = "0.8.3"
 authors = [{name = "Juan B Cabral & QuatroPe", email = "jbcabral@unc.edu.ar"}]
 readme = "README.md"
 license = {file = "LICENSE.txt"}

skcriteria/preprocessing/weighters.py

@@ -208,38 +208,39 @@ def _weight_matrix(self, matrix, **kwargs):
 
 
 def entropy_weights(matrix):
-    """Calculate the weights as the entropy of each criterion.
+    """Calculate the weights as the complement of the entropy of each \
+    criterion.
 
     It uses the underlying ``scipy.stats.entropy`` function which assumes that
     the values of the criteria are probabilities of a distribution.
 
-    This routine will normalize the criteria if they don’t sum to 1.
+    The logarithmic base to use is the number of rows/alternatives in the
+    matrix.
+
+    This routine will normalize the sum of the weights to 1.
 
     See Also
     --------
     scipy.stats.entropy :
         Calculate the entropy of a distribution for given probability values.
 
-    Examples
-    --------
-    >>> from skcriteria.preprocess import entropy_weights
-    >>> mtx = [[1, 2], [3, 4]]
-
-    >>> entropy_weights(mtx)
-    array([0.46906241, 0.53093759])
-
     """
-    entropy = scipy.stats.entropy(matrix, axis=0)
-    return entropy / np.sum(entropy)
+    base = len(matrix)
+    entropy = scipy.stats.entropy(matrix, base=base, axis=0)
+    entropy_divergence = 1 - entropy
+    return entropy_divergence / np.sum(entropy_divergence)
 
 
 class EntropyWeighter(SKCWeighterABC):
-    """Assigns the entropy of the criteria as weights.
+    """Assigns the complement of the entropy of the criteria as weights.
 
     It uses the underlying ``scipy.stats.entropy`` function which assumes that
     the values of the criteria are probabilities of a distribution.
 
-    This transformer will normalize the criteria if they don't sum to 1.
+    The logarithmic base to use is the number of rows/alternatives in the
+    matrix.
+
+    This transformer will normalize the sum of the weights to 1.
 
     See Also
     --------

tests/preprocessing/test_weighters.py

@@ -215,12 +215,12 @@ def test_EntropyWeighter_simple_matrix():
     expected = skcriteria.mkdm(
         matrix=[[1, 2], [4, 16]],
         objectives=[min, max],
-        weights=[0.589239, 0.410761],
+        weights=[0.358889, 0.641111],
     )
 
     weighter = EntropyWeighter()
-
     result = weighter.transform(dm)
+
     assert result.aequals(expected, atol=1e-5)
 
 
@@ -235,7 +235,7 @@ def test_EntropyWeighter(decision_matrix):
         min_objectives_proportion=0.5,
     )
 
-    entropy = scipy.stats.entropy(dm.matrix, axis=0)
+    entropy = 1 - scipy.stats.entropy(dm.matrix, base=10, axis=0)
 
     expected = skcriteria.mkdm(
         matrix=dm.matrix,
@@ -252,6 +252,28 @@ def test_EntropyWeighter(decision_matrix):
     assert result.equals(expected)
 
 
+def test_EntropyWeighter_less_predictable_more_weight():
+
+    dm = skcriteria.mkdm(
+        [
+            [1, 20, 300],
+            [1, 20, 400],
+            [1, 30, 500],
+            [1, 30, 600],
+            [1, 40, 700],
+            [1, 40, 800],
+        ],
+        objectives=[max, max, max],
+        criteria="C0 C1 C2".split(),
+    )
+
+    weighter = EntropyWeighter()
+    result = weighter.transform(dm)
+
+    assert result.weights["C0"] < result.weights["C1"]
+    assert result.weights["C1"] < result.weights["C2"]
+
+
 # =============================================================================
 # CRITIC
 # =============================================================================