Add tests for computation of plausible counterfactuals

tests/sklearn/test_sklearn_decisiontree.py

@@ -3,12 +3,95 @@
 sys.path.insert(0,'..')
 
 import numpy as np
+import random
+random.seed(424242)
 import sklearn
 from sklearn.datasets import load_iris, load_boston
-from sklearn.model_selection import train_test_split
+from sklearn.neighbors import KernelDensity
+from sklearn.mixture import GaussianMixture
+from sklearn.model_selection import GridSearchCV, train_test_split
 from sklearn.tree import DecisionTreeClassifier, DecisionTreeRegressor
+from sklearn.utils import shuffle
 
 from ceml.sklearn import generate_counterfactual
+from ceml.sklearn.decisiontree import decisiontree_generate_counterfactual
+from ceml.sklearn.plausibility import prepare_computation_of_plausible_counterfactuals
+
+
+def test_plausible_counterfactual():
+    # Load data
+    X, y = load_iris(return_X_y=True)
+
+    X, y = shuffle(X, y, random_state=42)
+    X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.33, random_state=4242)
+
+    # Choose target labels
+    y_test_target = []
+    labels = np.unique(y)
+    for i in range(X_test.shape[0]):
+        y_test_target.append(random.choice(list(filter(lambda l: l != y_test[i], labels))))
+    y_test_target = np.array(y_test_target)
+
+    # Reduce dimensionality
+    X_train_orig = np.copy(X_train)
+    X_test_orig = np.copy(X_test)
+    projection_matrix = None
+    projection_mean_sub = None
+
+    # Fit classifier
+    model = DecisionTreeClassifier(max_depth=7, random_state=42)
+    model.fit(X_train, y_train)
+
+    # For each class, fit density estimators
+    density_estimators = {}
+    kernel_density_estimators = {}
+    labels = np.unique(y)
+    for label in labels:
+        # Get all samples with the 'correct' label
+        idx = y_train == label
+        X_ = X_train[idx, :]
+
+        # Optimize hyperparameters
+        cv = GridSearchCV(estimator=KernelDensity(), param_grid={'bandwidth': np.arange(0.1, 10.0, 0.05)}, n_jobs=-1, cv=5)
+        cv.fit(X_)
+        bandwidth = cv.best_params_["bandwidth"]
+
+        cv = GridSearchCV(estimator=GaussianMixture(covariance_type='full'), param_grid={'n_components': range(2, 10)}, n_jobs=-1, cv=5)
+        cv.fit(X_)
+        n_components = cv.best_params_["n_components"]
+
+        # Build density estimators
+        kde = KernelDensity(bandwidth=bandwidth)
+        kde.fit(X_)
+
+        de = GaussianMixture(n_components=n_components, covariance_type='full', random_state=42)
+        de.fit(X_)
+
+        density_estimators[label] = de
+        kernel_density_estimators[label] = kde
+
+    # Build dictionary for ceml
+    plausibility_stuff = prepare_computation_of_plausible_counterfactuals(X_train_orig, y_train, density_estimators, projection_mean_sub, projection_matrix)
+
+    # Compute and plot counterfactual with vs. without density constraints
+    i = 0
+
+    x_orig = X_test[i,:]
+    x_orig_orig = X_test_orig[i,:]
+    y_orig = y_test[i]
+    y_target = y_test_target[i]
+
+    assert model.predict([x_orig]) == y_orig # Model already predicts target label!
+
+    # Compute plausible counterfactual
+    x_cf_plausible = decisiontree_generate_counterfactual(model, x_orig_orig, y_target, plausibility=plausibility_stuff)
+    print("Predictec label of plausible countrefactual: {0}".format(model.predict([x_cf_plausible])))
+    assert model.predict([x_cf_plausible]) == y_target
+
+    # Compute closest counterfactual     
+    plausibility_stuff["use_density_constraints"] = False   
+    x_cf = decisiontree_generate_counterfactual(model, x_orig_orig, y_target, plausibility=plausibility_stuff)
+    assert model.predict([x_cf]) == y_target
 
 
 def test_decisiontree_classifier():

tests/sklearn/test_sklearn_softmaxregression.py

@@ -4,14 +4,21 @@
 
 import numpy as np
 np.random.seed(42)
+import random
+random.seed(424242)
+from sklearn.utils import shuffle
 import pytest
 import sklearn
 from sklearn.datasets import load_iris
-from sklearn.model_selection import train_test_split
+from sklearn.neighbors import KernelDensity
+from sklearn.mixture import GaussianMixture
+from sklearn.model_selection import GridSearchCV, train_test_split
 from sklearn.linear_model import LogisticRegression
 from scipy.optimize import minimize
 
 from ceml.sklearn import generate_counterfactual
+from ceml.sklearn.softmaxregression import softmaxregression_generate_counterfactual
+from ceml.sklearn.plausibility import prepare_computation_of_plausible_counterfactuals
 from ceml.optim import Optimizer
 from ceml.backend.jax.costfunctions import LMadCost 
 
@@ -42,6 +49,82 @@ def __call__(self):
         return np.array(optimum["x"])
 
 
+def test_plausible_counterfactual():
+    # Load data
+    X, y = load_iris(return_X_y=True)
+
+    X, y = shuffle(X, y, random_state=42)
+    X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.33, random_state=4242)
+
+    # Choose target labels
+    y_test_target = []
+    labels = np.unique(y)
+    for i in range(X_test.shape[0]):
+        y_test_target.append(random.choice(list(filter(lambda l: l != y_test[i], labels))))
+    y_test_target = np.array(y_test_target)
+
+    # Reduce dimensionality
+    X_train_orig = np.copy(X_train)
+    X_test_orig = np.copy(X_test)
+    projection_matrix = None
+    projection_mean_sub = None
+
+    # Fit classifier
+    model = LogisticRegression(multi_class="multinomial", solver="lbfgs", random_state=42)
+    model.fit(X_train, y_train)
+
+    # For each class, fit density estimators
+    density_estimators = {}
+    kernel_density_estimators = {}
+    labels = np.unique(y)
+    for label in labels:
+        # Get all samples with the 'correct' label
+        idx = y_train == label
+        X_ = X_train[idx, :]
+
+        # Optimize hyperparameters
+        cv = GridSearchCV(estimator=KernelDensity(), param_grid={'bandwidth': np.arange(0.1, 10.0, 0.05)}, n_jobs=-1, cv=5)
+        cv.fit(X_)
+        bandwidth = cv.best_params_["bandwidth"]
+
+        cv = GridSearchCV(estimator=GaussianMixture(covariance_type='full'), param_grid={'n_components': range(2, 10)}, n_jobs=-1, cv=5)
+        cv.fit(X_)
+        n_components = cv.best_params_["n_components"]
+
+        # Build density estimators
+        kde = KernelDensity(bandwidth=bandwidth)
+        kde.fit(X_)
+
+        de = GaussianMixture(n_components=n_components, covariance_type='full', random_state=42)
+        de.fit(X_)
+
+        density_estimators[label] = de
+        kernel_density_estimators[label] = kde
+
+    # Build dictionary for ceml
+    plausibility_stuff = prepare_computation_of_plausible_counterfactuals(X_train_orig, y_train, density_estimators, projection_mean_sub, projection_matrix)
+
+    # Compute and plot counterfactual with vs. without density constraints
+    i = 0
+
+    x_orig = X_test[i,:]
+    x_orig_orig = X_test_orig[i,:]
+    y_orig = y_test[i]
+    y_target = y_test_target[i]
+
+    assert model.predict([x_orig]) == y_orig # Model already predicts target label!
+
+    # Compute plausible counterfactual
+    x_cf_plausible = softmaxregression_generate_counterfactual(model, x_orig_orig, y_target, plausibility=plausibility_stuff)
+    print("Predictec label of plausible countrefactual: {0}".format(model.predict([x_cf_plausible])))
+    assert model.predict([x_cf_plausible]) == y_target
+
+    # Compute closest counterfactual     
+    plausibility_stuff["use_density_constraints"] = False   
+    x_cf = softmaxregression_generate_counterfactual(model, x_orig_orig, y_target, plausibility=plausibility_stuff)
+    assert model.predict([x_cf]) == y_target
+
+
 def test_softmaxregression():
     # Load data
     X, y = load_iris(True)