Explanations with SHAP and fairness measures

adult.py

+import datetime
+
+from sklearn.ensemble import AdaBoostClassifier
+from sklearn.ensemble._forest import RandomForestClassifier
+from sklearn.ensemble._gb import GradientBoostingClassifier
+from sklearn.linear_model._logistic import LogisticRegression
+from sklearn.model_selection import train_test_split
+from sklearn.neural_network._multilayer_perceptron import MLPClassifier
+from sklearn.svm._classes import SVC
+from sklearn.tree import DecisionTreeClassifier
+
+from core import load_data, Model, evaluation, \
+    find_threshold, remove, ensemble_out, train_classifier
+from lime_global import fairness_eval
+
+
+import os
+import numpy as np
+
+os.environ['KMP_DUPLICATE_LIB_OK']='True'
+np.random.seed(10)
+
+def main(seed):
+    train_size = 0.8
+    max_features = 4
+    sample_size = 500
+    
+    algo = MLPClassifier
+    print(algo.__name__)
+    source_name = 'adult.data'
+    to_drop = [5,8,9] # sex, race, marital status
+    all_categorical_features = [1,3,5,6,7,8,9,13]
+    
+    data, labels, class_names, feature_names, categorical_names = load_data(source_name, all_categorical_features, delimiter=',')
+    train, test, labels_train, labels_test = train_test_split(data, labels, train_size=train_size, test_size=1-train_size, random_state=seed)
+    
+    print("Features")
+    for i in range(len(feature_names)):
+        f_name = feature_names[i]
+        print(i,f_name,end=' ')
+        if i in all_categorical_features:
+            print("[c]",end=' ')
+        if i in to_drop:
+            print("[s]",end=' ')
+        print('')    
+         
+    model, encoder = train_classifier(algo, train, test, labels_train, [], all_categorical_features)
+    original_model = Model([model],[encoder],[[]])
+    threshold_1 = find_threshold(original_model.prob(train), labels_train)
+    print("Original model OK")
+ 
+    fairness_eval(original_model, train, max_features, to_drop, feature_names, class_names, all_categorical_features, categorical_names, sample_size)
+    
+    ensemble = ensemble_out(algo,to_drop,train, test, labels_train, all_categorical_features)
+    threshold_2 = find_threshold(ensemble.prob(train), labels_train)
+    
+    print("ExpOut ensemble's model OK")
+    
+    accuracy_original = evaluation(original_model.prob(test), labels_test, threshold_1)
+    print("accuracy_original", accuracy_original)
+    accuracy_ensemble = evaluation(ensemble.prob(test), labels_test, threshold_2)
+    print("accuracy_ensemble", accuracy_ensemble)
+    
+    # debug
+    for i in range(len(ensemble.models)):
+        m = Model([ensemble.models[i]],[ensemble.encoders[i]],[ensemble.features_to_remove[i]])
+        print("accuracy_m",i,' ', evaluation(m.prob(test), labels_test, threshold_2), sep='')
+    
+    fairness_eval(ensemble, train, max_features, to_drop, feature_names, class_names, all_categorical_features, categorical_names, sample_size)
+
+    
+if __name__ == "__main__":
+    now = datetime.datetime.now()
+    print('adult mlp (100,100,)\n',now.year,'-', now.month,'-', now.day,',', now.hour,':', now.minute,':', now.second, sep='')
+    for i in range(5):
+        print("experiment i=",i)
+        main(i)
\ No newline at end of file

all_explainers.py

+ 
+from collections import Counter
+
+from anchor import anchor_tabular
+import lime_global
+import pandas as pd
+import numpy as np
+
+
+def fairness_eval(model, train, max_features, sensitive_features, feature_names, class_names, categorical_features, categorical_names):
+    
+    _, sp_obj = lime_global.features_contributions(model.prob, train, feature_names, max_features, class_names, categorical_features, categorical_names)
+    
+    a = Counter()
+    for i in sp_obj.V:
+        exp = sp_obj.explanations[i]
+        a1 = Counter(dict(exp.local_exp[1]))
+        a.update(a1)
+    
+    ans_data = []
+    for key in a:
+        ans_data1 = [feature_names[key],a[key]]
+        ans_data.append(ans_data1)
+        
+    df = pd.DataFrame(ans_data, columns = ["Feature", "Contribution"])
+    print(df.iloc[(-np.abs(df['Contribution'].values)).argsort()])
+    
+    
+    indices = sp_obj.indices
+    a_explainer = anchor_tabular.AnchorTabularExplainer(class_names,feature_names,train,categorical_names=categorical_names)
+    
+    non_empty_anchors = 0
+    counter = Counter()
+    for i in indices:
+        exp = a_explainer.explain_instance(train[i], model.predict, threshold=0.95)
+        print(i,'%.2f' % exp.precision(),' %.2f' % exp.coverage(), ' (class %s)' % exp.exp_map['prediction'], '%s' % (' AND '.join(exp.names())))
+        features = exp.exp_map['feature']
+        if len(features) > 0:
+            a1 = Counter(features)
+            non_empty_anchors += 1
+            counter.update(a1)
+    
+    ans_data = []
+    for key, value in sorted(counter.items(), key=lambda x: x[1], reverse=True):
+        ans_data1 = [feature_names[key],value/non_empty_anchors]
+        ans_data.append(ans_data1)
+        
+    df = pd.DataFrame(ans_data, columns = ["Feature", "Frequency"])
+    print(df.iloc[(-np.abs(df['Frequency'].values)).argsort()])
+    
+    

anchor_global.py

@@ -45,7 +45,7 @@ def fairness_eval(model, train, max_features, sensitive_features, feature_names,
     print(df.iloc[(-np.abs(df['Frequency'].values)).argsort()])
     
     
-    return is_fair, ans_data    
+    return is_fair, ans_data, a_explainers    
 
     
 
@@ -54,8 +54,8 @@ def fairness_eval2(model, train, max_features, sensitive_features, feature_names
 #     _, sp_obj = lime_global.features_contributions(model.prob, train, feature_names, max_features, class_names, categorical_features, categorical_names, sample_size)
     
 #     indices = sp_obj.indices
-#     indices = random.choices(range(len(train)), k=500)
-    indices = list(range(800))
+    indices = random.choices(range(len(train)), k=sample_size)
+#     indices = list(range(800))
     a_explainers = anchor_tabular.AnchorTabularExplainer(class_names,feature_names,train,categorical_names=categorical_names)
     
     anchors_pos = []
@@ -167,7 +167,7 @@ def pareto_set(anchors):
       
     return optimal_anchors
 
-def multi_pareto_set(anchors,n=50):
+def multi_pareto_set(anchors,n=1):
     
     current_anchors = anchors.copy()
     optimal_anchors = []

bank.py

+import datetime
+
+from sklearn.ensemble import AdaBoostClassifier
+from sklearn.ensemble._forest import RandomForestClassifier
+from sklearn.linear_model._logistic import LogisticRegression
+from sklearn.model_selection import train_test_split
+from sklearn.neural_network._multilayer_perceptron import MLPClassifier
+from sklearn.svm._classes import SVC
+from sklearn.tree import DecisionTreeClassifier
+
+from anchor_global import fairness_eval
+from core import load_data, Model, evaluation, \
+    find_threshold, remove, ensemble_out, train_classifier
+
+
+def main(seed):
+    train_size = 0.8
+    max_features = 10
+    
+    algo = SVC
+    print(algo.__name__)
+    source_name = 'bank.data'
+    to_drop = [0,2] # age, marital 
+    all_categorical_features = [1,2,3,4,6,7,8,10,15] 
+    
+    data, labels, class_names, feature_names, categorical_names = load_data(source_name, all_categorical_features,delimiter=";")
+    train, test, labels_train, labels_test = train_test_split(data, labels, train_size=train_size, test_size=1-train_size, random_state=seed)
+    
+     
+    model, encoder = train_classifier(algo, train, test, labels_train, [], all_categorical_features)
+    original_model = Model([model],[encoder],[[]])
+#     threshold_1 = find_threshold(original_model.prob(train), labels_train)
+    print("Original model OK")
+
+    fairness_eval(original_model, train, max_features, to_drop, feature_names, class_names, all_categorical_features, categorical_names)
+    
+    ensemble = ensemble_out(algo,to_drop,train, test, labels_train, all_categorical_features)
+#     threshold_2 = find_threshold(ensemble.prob(train), labels_train)
+    
+    print("ExpOut ensemble's model OK")
+    
+#     accuracy_original = evaluation(original_model.prob(test), labels_test, threshold_1)
+#     print("accuracy_original", accuracy_original)
+#     accuracy_ensemble = evaluation(ensemble.prob(test), labels_test, threshold_2)
+#     print("accuracy_ensemble", accuracy_ensemble)
+    
+    fairness_eval(ensemble, train, max_features, to_drop, feature_names, class_names, all_categorical_features, categorical_names)
+
+    
+if __name__ == "__main__":
+    now = datetime.datetime.now()
+    print('bank\n',now.year,'-', now.month,'-', now.day,',', now.hour,':', now.minute,':', now.second, sep='')
+    for i in range(1):
+        print("experiment i=",i)
+        main(i)
\ No newline at end of file

batch_experiments.py

+import datetime
+
+from sklearn.ensemble import AdaBoostClassifier
+from sklearn.ensemble._forest import RandomForestClassifier
+from sklearn.linear_model._logistic import LogisticRegression
+from sklearn.model_selection import train_test_split
+from sklearn.neural_network._multilayer_perceptron import MLPClassifier
+from sklearn.tree import DecisionTreeClassifier
+
+from anchor_global import features_contributions
+from core import load_data, train_classifier, Model, evaluation, \
+    find_threshold, remove, ensemble_out, random_state
+
+def main():
+    train_size = 0.8
+    test_size = 0.2
+    
+    algo = MLPClassifier
+    print(algo.__name__)
+    source_name = 'german.data'
+    feature_names = ['existingchecking', 'duration', 'credithistory', 'purpose', 'creditamount', 'savings', 'employmentsince', 'installmentrate', 'statussex', 'otherdebtors', 'residencesince', 'property', 'age', 'otherinstallmentplans', 'housing', 'existingcredits', 'job', 'peopleliable', 'telephone', 'foreignworker', 'classification']
+    to_drop = [8,18,19]
+    all_categorical_features = [0,2,3,5,6,8,9,11,13,14,16,18,19]
+    
+    data, labels, class_names, feature_names, categorical_names = load_data(source_name, feature_names, all_categorical_features)
+    
+    for i in range(30):
+    
+        train, test, labels_train, labels_test = train_test_split(data, labels, train_size=train_size, test_size=test_size, random_state=i)
+        
+        model, encoder = train_classifier(algo, train, labels_train, [], all_categorical_features)
+        original_model = Model([model],[encoder],[[]])
+        
+        threshold_1 = find_threshold(original_model.prob(train), labels_train)
+        accuracy1 = evaluation(original_model.prob(test), labels_test, threshold_1)
+        
+        ensemble = ensemble_out(algo,to_drop,train,labels_train, all_categorical_features)
+
+        threshold_2 = find_threshold(ensemble.prob(train), labels_train)
+        accuracy2 = evaluation(ensemble.prob(test), labels_test, threshold_2)
+    
+        print(i,accuracy1,accuracy2)
+    
+
+    
+if __name__ == "__main__":
+    now = datetime.datetime.now()
+    print(now.year,'-', now.month,'-', now.day,',', now.hour,':', now.minute,':', now.second)
+    main()
\ No newline at end of file

compas.py

+import datetime
+
+from sklearn.ensemble import AdaBoostClassifier
+from sklearn.ensemble._forest import RandomForestClassifier
+from sklearn.linear_model._logistic import LogisticRegression
+from sklearn.model_selection import train_test_split
+from sklearn.neural_network._multilayer_perceptron import MLPClassifier
+from sklearn.tree import DecisionTreeClassifier
+
+from anchor_global import fairness_eval
+from core import load_data, Model, evaluation, \
+    find_threshold, remove, ensemble_out, train_classifier
+from sklearn.svm import SVC
+
+def main(seed):
+    train_size = 0.8
+    max_features = 10
+    
+    algo = SVC
+    print(algo.__name__)
+    source_name = 'compas.data'
+    to_drop = [4,7,8,9] #  age, age_cat, sex, race
+    all_categorical_features = [5,6,7,8,9]
+    target_features = ["juv_fel_count", "juv_misd_count", "juv_other_count", "priors_count","age", "c_charge_degree", "c_charge_desc", "age_cat", "sex", "race","is_recid"]
+    
+    data, labels, class_names, feature_names, categorical_names = load_data(source_name, all_categorical_features,delimiter=",", target_features=target_features)
+    train, test, labels_train, labels_test = train_test_split(data, labels, train_size=train_size, test_size=1-train_size, random_state=seed)
+    
+     
+    model, encoder = train_classifier(algo, train, test, labels_train, [], all_categorical_features)
+    original_model = Model([model],[encoder],[[]])
+#     threshold_1 = find_threshold(original_model.prob(train), labels_train)
+    print("Original model OK")
+
+    fairness_eval(original_model, train, max_features, to_drop, feature_names, class_names, all_categorical_features, categorical_names)
+    
+    ensemble = ensemble_out(algo,to_drop,train, test, labels_train, all_categorical_features)
+#     threshold_2 = find_threshold(ensemble.prob(train), labels_train)
+    
+    print("ExpOut ensemble's model OK")
+    fairness_eval(ensemble, train, max_features, to_drop, feature_names, class_names, all_categorical_features, categorical_names)
+#     
+#     accuracy_original = evaluation(original_model.prob(test), labels_test, threshold_1)
+#     print("accuracy_original", accuracy_original)
+#     accuracy_ensemble = evaluation(ensemble.prob(test), labels_test, threshold_2)
+#     print("accuracy_ensemble", accuracy_ensemble)
+    
+
+    
+if __name__ == "__main__":
+    now = datetime.datetime.now()
+    print('compas\n',now.year,'-', now.month,'-', now.day,',', now.hour,':', now.minute,':', now.second, sep='')
+    for i in range(1):
+        print("experiment i=",i)
+        main(i)
\ No newline at end of file

core.py

@@ -3,10 +3,17 @@ Implements the main procedures to build fairer ensembles, e.g. feature drop out,
 """
 import os
 
+from aif360 import metrics
+from aif360.sklearn.metrics.metrics import statistical_parity_difference, \
+    equal_opportunity_difference, average_odds_difference, \
+    disparate_impact_ratio, difference, generalized_fpr, specificity_score
+from aif360.sklearn.utils import check_groups
 from imblearn.over_sampling import SMOTE
 from sklearn.compose import ColumnTransformer
 from sklearn.metrics import accuracy_score
 from sklearn.metrics import f1_score
+from sklearn.metrics._classification import confusion_matrix
+from sklearn.model_selection._split import train_test_split
 from sklearn.preprocessing._encoders import OneHotEncoder
 from sklearn.preprocessing._label import LabelEncoder
 
@@ -15,8 +22,125 @@ import pandas as pd
 
 
 os.environ['KMP_DUPLICATE_LIB_OK']='True'
-# np.random.seed(10)
 
+
+class FixOut:
+    
+    def __init__(self,source_name, sep, train_size, to_drop, all_categorical_features, algo, exp, max_features, sampling_size, seed=None, threshold=None):
+            
+        self.exp = exp
+        self.algo = algo
+        self.seed = seed
+        self.threshold = threshold
+
+        self.data, self.labels, self.class_names, self.feature_names, self.categorical_names = load_data(source_name, all_categorical_features, delimiter=sep)
+        self.train, self.test, self.labels_train, self.labels_test = train_test_split(self.data, self.labels, train_size=train_size, random_state=self.seed)
+        
+        self.sensitive_f = to_drop
+        self.all_categorical_f =  all_categorical_features
+        self.max_features = max_features
+        self.sampling_size = sampling_size
+        
+#         print("Features")
+#         for i in range(len(self.feature_names)):
+#             f_name = self.feature_names[i]
+#             print(i,f_name,end=' ')
+#             if i in all_categorical_features:
+#                 print("[c]",end=' ')
+#             if i in to_drop:
+#                 print("[s]",end=' ')
+#             print('')
+    
+    
+    def is_fair(self):
+        
+        model, encoder = train_classifier(self.algo, self.train, self.test, self.labels_train, [], self.all_categorical_f, self.seed)
+        self.original_model = Model([model],[encoder],[[]])
+        
+        accuracy, _ = evaluation(self.original_model.prob(self.test), self.labels_test)
+    
+        actual_sensitive, is_fair_flag ,ans_data, explainer = self.exp(self.original_model, self.train, self.max_features, self.sensitive_f, self.feature_names, self.class_names, self.all_categorical_f, self.categorical_names, self.sampling_size, self.threshold)
+         
+        return actual_sensitive, is_fair_flag, ans_data, accuracy
+    
+    def ensemble_out(self, actual_sensitive):
+        
+        """
+        Implements ENSEMBLE_Out 
+        
+        Parameters
+            algo: class of a classification algorithm
+            to_drop: list of features that must be dropped 
+            train: X
+            labels_train: y
+            all_categorical_features: list of indices of categorical features
+            
+        """
+    
+        models, encoders, features_to_remove = [], [], []
+        
+        for i in actual_sensitive:
+            remove_features = [i]
+            categorical_features = remove(remove_features, self.all_categorical_f)
+            model, encoder = train_classifier(self.algo, self.train, self.test, self.labels_train, remove_features, categorical_features, self.seed)
+            models.append(model)
+            encoders.append(encoder)
+            features_to_remove.append(remove_features)
+        
+        categorical_features4 = remove(actual_sensitive, self.all_categorical_f)
+        model4, encoder4 = train_classifier(self.algo, self.train, self.test, self.labels_train, actual_sensitive, categorical_features4, self.seed)
+        models.append(model4)
+        encoders.append(encoder4)
+        features_to_remove.append(actual_sensitive)
+        
+        self.ensemble = Model(models,encoders,features_to_remove)
+        
+        accuracy, _ = evaluation(self.ensemble.prob(self.test), self.labels_test)
+    
+#         _, is_fair_flag ,ans_data, explainer = self.exp(self.ensemble, self.train, self.max_features, actual_sensitive, self.feature_names, self.class_names, self.all_categorical_f, self.categorical_names, self.sampling_size, self.threshold)
+#          
+#         return is_fair_flag, ans_data, accuracy
+        return False, None, accuracy
+        
+
+class Model:
+    """Class for ensemble models
+    
+    Saves a list of trained classifiers and their respective encoders and deleted features
+    
+    """
+    def __init__(self,models,encoders,features_to_remove):
+        self.models = models
+        self.encoders = encoders
+        self.features_to_remove = features_to_remove
+        
+    """
+    Args:
+        models: a list of trained classifiers
+        encoders: a list of encoders (1st encoder for the 1st model)  
+        features_to_remove: a list of lists of indices to be removed in order to use models (feature removal mapping)  
+    """
+           
+    def prob(self,X):
+        """
+        Returns probability for each class label.
+        """
+        
+        probs = []
+        n_models = len(self.models)
+        
+        for i in range(n_models):
+            model = self.models[i]
+            encoder = self.encoders[i]
+            to_remove = self.features_to_remove[i]
+            comp = model.predict_proba(encoder.transform(np.delete(X, to_remove, axis=1))).astype(float)
+            probs.append(comp)
+           
+        res = sum(probs)/n_models
+        
+        return res
+
+        
 def load_data(source_name, categorical_features, feature_names=None, delimiter=' ', target_features=None):
     """
     Loads data from a text file source
@@ -32,9 +156,15 @@ def load_data(source_name, categorical_features, feature_names=None, delimiter='
         data.dropna(subset=target_features, inplace=True)
         
     current_feature_names = data.columns.values.tolist()
+    current_feature_names = [f_name.replace(" ", "") for f_name in current_feature_names]
     data = data.replace(np.nan, '', regex=True)
+#     data = data.replace('', 0, regex=True)
 #     data.iloc[categorical_features] = data.iloc[categorical_features].astype(str)
         
+    for i in get_numerical_features_indexes(len(current_feature_names), categorical_features):
+        data.iloc[:,i] = data.iloc[:,i].replace('', 0, regex=False)
+        data.iloc[:,i] = data.iloc[:,i].replace('?', 0, regex=False)
+
     data = data.to_numpy()
     
     labels = data[:,-1]
@@ -48,15 +178,18 @@ def load_data(source_name, categorical_features, feature_names=None, delimiter='
     categorical_names = {}
     for feature in categorical_features:
         le = LabelEncoder()
-        le.fit(data[:, feature])
-        data[:, feature] = le.transform(data[:, feature])
+        column = data[:, feature].astype(str)
+        le.fit(column)
+        data[:, feature] = le.transform(column)
         categorical_names[feature] = le.classes_
+        
+        
     data = data.astype(float)
     
     return data, labels, class_names, current_feature_names, categorical_names
 
 
-def train_classifier(algo, train, test, train_labels, remove_features, categorical_features):
+def train_classifier(algo, train, test, train_labels, remove_features, categorical_features, seed):
     
     train = np.delete(train, remove_features, axis = 1)
     test = np.delete(test, remove_features, axis = 1)
@@ -65,96 +198,28 @@ def train_classifier(algo, train, test, train_labels, remove_features, categoric
         [('one_hot_encoder', OneHotEncoder(categories='auto'),  categorical_features)],   
         remainder='passthrough'
     )
-    end = encoder.fit(np.concatenate([train, test]))
-    encoded_train = end.transform(train)
+    encoder.fit(np.concatenate([train, test]))
     
-    sm = SMOTE(sampling_strategy='auto')
-    train_res, labels_train_res = sm.fit_sample(encoded_train, train_labels)
+    sm = SMOTE(sampling_strategy='auto', random_state=seed)
+    train_res, labels_train_res = sm.fit_sample(train, train_labels)
+
+    encoder.fit(np.concatenate([train_res, test]))
+    encoded_train_res = encoder.transform(train_res)
           
+#     model = algo(hidden_layer_sizes=(100,100,),random_state=seed)
     model = algo()
-    model.fit(train_res, labels_train_res)
+    model.fit(encoded_train_res, labels_train_res)
     
     return model, encoder
 
-def ensemble_out(algo,to_drop,train,test,labels_train, all_categorical_features):
-    """
-    Implements ENSEMBLE_Out 
-    
-    Parameters
-        algo: class of a classification algorithm
-        to_drop: list of features that must be dropped 
-        train: X
-        labels_train: y
-        all_categorical_features: list of indices of categorical features
-        
-    """
-
-    models, encoders, features_to_remove = [], [], []
+def get_numerical_features_indexes(num_features, categorical_features):
     
-    for i in to_drop:
-        remove_features = [i]
-        categorical_features = remove(remove_features, all_categorical_features)
-        model, encoder = train_classifier(algo, train, test, labels_train, remove_features, categorical_features)
-        models.append(model)
-        encoders.append(encoder)
-        features_to_remove.append(remove_features)
+    res = []
+    for i in range(num_features):
+        if i not in categorical_features:
+            res.append(i)           
+    return res 
     
-    categorical_features4 = remove(to_drop, all_categorical_features)
-    model4, encoder4 = train_classifier(algo, train, test, labels_train, to_drop, categorical_features4)
-    models.append(model4)
-    encoders.append(encoder4)
-    features_to_remove.append(to_drop)
-    
-    return Model(models,encoders,features_to_remove)
-
-class Model:
-    """Class for ensemble models
-    
-    Saves a list of trained classifiers and their respective encoders and deleted features
-    
-    """
-    def __init__(self,models,encoders,features_to_remove):
-        self.models = models
-        self.encoders = encoders
-        self.features_to_remove = features_to_remove
-        
-    """
-    Args:
-        models: a list of trained classifiers
-        encoders: a list of encoders (1st encoder for the 1st model)  
-        features_to_remove: a list of lists of indices to be removed in order to use models (feature removal mapping)  
-    """
-           
-    def prob(self,X):
-        """
-        Returns probability for each class label.
-        """
-        
-        probs = []
-        n_models = len(self.models)
-        
-        for i in range(n_models):
-            model = self.models[i]
-            encoder = self.encoders[i]
-            to_remove = self.features_to_remove[i]
-#             print("model ",i)
-            comp = model.predict_proba(encoder.transform(np.delete(X, to_remove, axis=1))).astype(float)
-            probs.append(comp)
-           
-        res = sum(probs)/n_models
-