Change logic in change factor calculation

The current change tries to minimize the distance and increase the counterfactual score (therefore, lower distances and higher prediction scores are preferred)

cfnow/_fine_tune.py

@@ -33,10 +33,25 @@ def _calculate_change_factor(c_cf, changes_back_factual, feat_distances, changes
     # Now, calculate the probability for each different index
     cf_back_factual_probs = mp1c(cf_back_factual)
 
+    prediction_dif = (c_cf_c - cf_back_factual_probs)
+
     # Calculate how much a unitary change cause a change in probability
-    change_factor_feat = (c_cf_c - cf_back_factual_probs) / feat_distances[changes_back_original_idxs]
+    # The probability should be as negative as possible
+    # The distance (objective function) should be as close to zero as possible (as it cannot be negative)
+    # For a positive probability, we just multiply the prediction change by the distance
+    # For a negative probability, we just divide the prediction change by the distance
+    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)
 
-    return change_factor_feat
+    return np.array(change_factor_feat)
 
 
 def _generate_change_vectors(factual, factual_np, c_cf, _feat_idx_to_type, tabu_list,

cfnow/_obj_functions.py

@@ -2,6 +2,10 @@
 In this module, we can specify the objective functions which the CF generator can use to optimize.
 The optimizer will try to get the lowest value possible (minimization) of the objective function under the
 condition the CF still a CF.
+
+The requirements for a valid objective function are:
+    - Lower values should be better.
+    - It should not return negative values or zero. (obj > 0)
 """
 import numpy as np
 

tests/test__fine_tune.py

@@ -46,7 +46,38 @@ 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, 0.125, 0.25, 0.5], list([round(n, 3) for n in change_factor_feat]))
+        self.assertListEqual([0.1, 0.08, 0.04, 0.02], list([round(n, 3) for n in change_factor_feat]))
+        self.assertIsInstance(change_factor_feat, np.ndarray)
+
+    def test__calculate_change_factor_negative(self):
+        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, 0.4, 0.2])
+        changes_back_original_idxs = [0, 1, 2, 3]
+        mp1c = MagicMock()
+        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), 0.04, 0.02], 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):
+        # 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])
+        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.assertIsInstance(change_factor_feat, np.ndarray)
 
     def test__generate_change_vectors_all_num_no_tabu(self):