Modify change factor calculation to favor large distances changes

Since it's more benefitial to allow big changes that cost low positive or high negative prediction changes, this commit makes this modification. Where small positive changes that greatly change the distance are favored and large negative changes that largely change the distance are also favored. The change in distance must be interpreted as how much closer the CF is from a factual, therefore, larger changes are better.

cfnow/_fine_tune.py

@@ -43,13 +43,12 @@ def _calculate_change_factor(c_cf, changes_back_factual, feat_distances, changes
     change_factor_feat = []
     for f_pc, f_d in zip(prediction_dif, feat_distances[changes_back_original_idxs]):
         if f_pc >= 0:
-            change_factor_feat.append(f_pc*f_d)
-        else:
-            # Avoid division by zero
             if f_d == 0:
                 change_factor_feat.append(0)
             else:
                 change_factor_feat.append(f_pc/f_d)
+        else:
+            change_factor_feat.append(f_pc*f_d)
 
     return np.array(change_factor_feat)
 

tests/test__fine_tune.py

@@ -46,7 +46,7 @@ def test__calculate_change_factor(self):
         change_factor_feat = _calculate_change_factor(c_cf, changes_back_factual, feat_distances,
                                                       changes_back_original_idxs,  mp1c, c_cf_c)
 
-        self.assertListEqual([0.1, 0.08, 0.04, 0.02], list([round(n, 3) for n in change_factor_feat]))
+        self.assertListEqual([0.1, 0.125, 0.25, 0.5], list([round(n, 3) for n in change_factor_feat]))
         self.assertIsInstance(change_factor_feat, np.ndarray)
 
     def test__calculate_change_factor_negative(self):
@@ -61,23 +61,23 @@ def test__calculate_change_factor_negative(self):
         change_factor_feat = _calculate_change_factor(c_cf, changes_back_factual, feat_distances,
                                                       changes_back_original_idxs,  mp1c, c_cf_c)
 
-        self.assertListEqual([-0.1, round(-0.1/0.8, 3), 0.04, 0.02], list([round(n, 3) for n in change_factor_feat]))
+        self.assertListEqual([-0.1, -0.08, 0.25, 0.5], list([round(n, 3) for n in change_factor_feat]))
         self.assertIsInstance(change_factor_feat, np.ndarray)
 
-    def test__calculate_change_factor_negative_zero_distance(self):
+    def test__calculate_change_factor_positive_zero_distance(self):
         # This should not happen as the objective function should not return zero values
         c_cf = np.array([1, 1, 1, 1])
         changes_back_factual = np.array([[0.1, 0., 0., 0.], [0., 0.1, 0., 0.], [0., 0., 0.1, 0.], [0., 0., 0., 0.1]])
         feat_distances = np.array([1, 0.8, 0, 0.2])
         changes_back_original_idxs = [0, 1, 2, 3]
         mp1c = MagicMock()
-        mp1c.return_value = np.array([0.7, 0.7, 0.7, 0.5])
+        mp1c.return_value = np.array([0.7, 0.7, 0.5, 0.5])
         c_cf_c = 0.6
 
         change_factor_feat = _calculate_change_factor(c_cf, changes_back_factual, feat_distances,
                                                       changes_back_original_idxs,  mp1c, c_cf_c)
 
-        self.assertListEqual([-0.1, round(-0.1/0.8, 3), 0, 0.02], list([round(n, 3) for n in change_factor_feat]))
+        self.assertListEqual([-0.1, -0.08, 0, 0.5], list([round(n, 3) for n in change_factor_feat]))
         self.assertIsInstance(change_factor_feat, np.ndarray)
 
     def test__generate_change_vectors_all_num_no_tabu(self):