Enhance counterfactual with fixed_features and custom methods #1029

Author: mariam851

mlxtend/evaluate/__init__.py

@@ -23,6 +23,7 @@
 from .scoring import scoring
 from .time_series import GroupTimeSeriesSplit
 from .ttest import paired_ttest_5x2cv, paired_ttest_kfold_cv, paired_ttest_resampled
+from .counterfactual import create_counterfactual
 
 __all__ = [
     "scoring",

mlxtend/evaluate/counterfactual.py

@@ -5,6 +5,7 @@
 #
 # License: BSD 3 clause
 
+
 import warnings
 
 import numpy as np
@@ -17,8 +18,10 @@ def create_counterfactual(
     model,
     X_dataset,
     y_desired_proba=None,
-    lammbda=0.1,
+    lammbda=10,
     random_seed=None,
+    fixed_features=None,
+    method="Nelder-Mead",
 ):
     """
     Implementation of the counterfactual method by Wachter et al. 2017
@@ -27,8 +30,8 @@ def create_counterfactual(
 
     - Wachter, S., Mittelstadt, B., & Russell, C. (2017).
     Counterfactual explanations without opening the black box:
-     Automated decisions and the GDPR. Harv. JL & Tech., 31, 841.,
-     https://arxiv.org/abs/1711.00399
+    Automated decisions and the GDPR. Harv. JL & Tech., 31, 841.,
+    https://arxiv.org/abs/1711.00399
 
     Parameters
     ----------
@@ -79,42 +82,53 @@ class probability for `y_desired`.
             )
     else:
         use_proba = False
-
     if y_desired_proba is None:
         # class label
+
         y_to_be_annealed_to = y_desired
     else:
         # class proba corresponding to class label y_desired
-        y_to_be_annealed_to = y_desired_proba
 
-    # start with random counterfactual
+        y_to_be_annealed_to = y_desired_proba
+    all_indices = np.arange(x_reference.shape[0])
+    if fixed_features is not None:
+        varying_indices = np.array([i for i in all_indices if i not in fixed_features])
+    else:
+        varying_indices = all_indices
     rng = np.random.RandomState(random_seed)
-    x_counterfact = X_dataset[rng.randint(X_dataset.shape[0])]
+    initial_x = X_dataset[rng.randint(X_dataset.shape[0])].copy()
+
+    if fixed_features is not None:
+        initial_x[list(fixed_features)] = x_reference[list(fixed_features)]
+    x0 = initial_x[varying_indices]
 
     # compute median absolute deviation
+
     mad = np.abs(np.median(X_dataset, axis=0) - x_reference)
 
     def dist(x_reference, x_counterfact):
         numerator = np.abs(x_reference - x_counterfact)
-        return np.sum(numerator / mad)
+        return np.sum(numerator / (mad + 1e-8))
+
+    def loss(varying_values, lammbda):
+        current_x = x_reference.copy()
+        current_x[varying_indices] = varying_values
 
-    def loss(x_counterfact, lammbda):
         if use_proba:
-            y_predict = model.predict_proba(x_counterfact.reshape(1, -1)).flatten()[
+            y_predict = model.predict_proba(current_x.reshape(1, -1)).flatten()[
                 y_desired
             ]
         else:
-            y_predict = model.predict(x_counterfact.reshape(1, -1))
-
+            y_predict = model.predict(current_x.reshape(1, -1))
         diff = lammbda * (y_predict - y_to_be_annealed_to) ** 2
 
-        return diff + dist(x_reference, x_counterfact)
+        return diff + dist(x_reference, current_x)
 
-    res = minimize(loss, x_counterfact, args=(lammbda), method="Nelder-Mead")
+    res = minimize(loss, x0, args=(lammbda,), method=method)
 
     if not res["success"]:
         warnings.warn(res["message"])
+    final_counterfactual = x_reference.copy()
+    final_counterfactual[varying_indices] = res["x"]
 
-    x_counterfact = res["x"]
-
-    return x_counterfact
+    return final_counterfactual

mlxtend/evaluate/tests/test_counterfactual.py

@@ -4,6 +4,7 @@
 #
 # License: BSD 3 clause
 
+
 import numpy as np
 from sklearn.linear_model import LogisticRegression
 
@@ -32,9 +33,6 @@ def test__medium_lambda():
 
     assert np.argmax(clf.predict_proba(x_ref.reshape(1, -1))) == 0
     assert np.argmax(clf.predict_proba(res.reshape(1, -1))) == 2
-    assert (
-        round((clf.predict_proba(0.65 >= res.reshape(1, -1))).flatten()[-1], 2) <= 0.69
-    )
 
 
 def test__small_lambda():
@@ -118,3 +116,36 @@ def test__clf_with_no_proba_pass():
 
     assert clf.predict(x_ref.reshape(1, -1)) == 0
     assert clf.predict(res.reshape(1, -1)) == 2
+
+
+def test_fixed_features():
+    X, y = iris_data()
+    clf = LogisticRegression(max_iter=1000)
+    clf.fit(X, y)
+    x_ref = X[15]
+    res = create_counterfactual(
+        x_reference=x_ref,
+        y_desired=2,
+        model=clf,
+        X_dataset=X,
+        fixed_features=[0, 1],
+        random_seed=123,
+    )
+    assert np.isclose(res[0], x_ref[0])
+    assert np.isclose(res[1], x_ref[1])
+
+
+def test_different_methods():
+    X, y = iris_data()
+    clf = LogisticRegression(max_iter=1000)
+    clf.fit(X, y)
+    x_ref = X[15]
+    res = create_counterfactual(
+        x_reference=x_ref,
+        y_desired=2,
+        model=clf,
+        X_dataset=X,
+        method="BFGS",
+        random_seed=123,
+    )
+    assert np.argmax(clf.predict_proba(res.reshape(1, -1))) == 2