fairml final

fairness_analysis/fairML_preprocessing.py

@@ -1,129 +1,90 @@
 import pandas as pd
-import numpy as np
 from sklearn.model_selection import train_test_split
 from sklearn.preprocessing import MinMaxScaler, OneHotEncoder
 from sklearn.compose import ColumnTransformer
 from sklearn.pipeline import Pipeline
 from sklearn.impute import SimpleImputer
-from sklearn.linear_model import LogisticRegression
-from imblearn.over_sampling import RandomOverSampler
 from sklearn.base import BaseEstimator, TransformerMixin
-from fairml import audit_model
-import matplotlib.pyplot as plt
-from fairml import plot_dependencies
+from imblearn.over_sampling import RandomOverSampler
+from sklearn.linear_model import LogisticRegression
+from fairml import audit_model, plot_dependencies
 from imblearn.pipeline import Pipeline as ImbPipeline
+import matplotlib.pyplot as plt
+
 
-# custom transformer to drop columns
-class DropColumns(BaseEstimator, TransformerMixin):
-    def __init__(self, columns):
-        self.columns = columns
+class DataFrameEnsurer(BaseEstimator, TransformerMixin):
+    """Transforms an array back to a dataframe, ensuring proper column names after transformations."""
+    def __init__(self, preprocessor):
+        self.preprocessor = preprocessor
+        self.feature_names = None
 
     def fit(self, X, y=None):
+        self.feature_names = self.preprocessor.get_feature_names_out()
         return self
 
     def transform(self, X):
-        return X.drop(self.columns, axis=1)
+        if self.feature_names is None:
+            raise Exception("The transformer is not yet fitted with feature names.")
+        return pd.DataFrame(X, columns=self.feature_names)
 
-# load data
-def load_data(filepath):
+# Load and preprocess data
+def preprocess_data(filepath):
     df = pd.read_csv(filepath, delimiter=';')
     df['y'] = df['y'].map({'yes': 1, 'no': 0})
-    return df
-
-class DataFrameTransformer(BaseEstimator, TransformerMixin):
-    def __init__(self, columns):
-        self.columns = columns  # columns should be a list of column names
-
-    def fit(self, X, y=None):
-        return self
-
-    def transform(self, X):
-        return pd.DataFrame(X, columns=self.columns)
+    df = df.drop(['contact', 'poutcome', 'duration'], axis=1) # drop columns that are not informative
 
-def preprocess_features(df):
+    # categorical and numerical columns
     categorical_cols = [col for col in df.columns if df[col].dtype == 'object']
-    numerical_cols = [col for col in df.select_dtypes(include=['int64', 'float64']).columns if col not in ['y']]
-
-    numerical_transformer = Pipeline(steps=[
-        ('imputer', SimpleImputer(strategy='mean')),
-        ('scaler', MinMaxScaler()),
-        ('to_df', DataFrameTransformer(numerical_cols))  # Convert back to DataFrame
-    ])
-
-    categorical_transformer = Pipeline(steps=[
-        ('imputer', SimpleImputer(strategy='constant', fill_value='missing')),
-        ('onehot', OneHotEncoder(handle_unknown='ignore')),
-        ('to_df', DataFrameTransformer(categorical_cols))  # Convert back to DataFrame
-    ])
-
-    # Combine using ColumnTransformer
-    preprocessor = ColumnTransformer(
-        transformers=[
-            ('num', numerical_transformer, numerical_cols),
-            ('cat', categorical_transformer, categorical_cols)
-        ], remainder='passthrough')  # 'passthrough' to keep columns as DataFrame if untouched
+    numerical_cols = [col for col in df.columns if df[col].dtype in ['int64', 'float64'] and col != 'y']
 
-    return preprocessor
-
-""""
-# Preprocess features
-def preprocess_features(df):
-    categorical_cols = [col for col in df.columns if df[col].dtype == 'object']
-    numerical_cols = [col for col in df.select_dtypes(include=['int64', 'float64']).columns if col not in ['y']]
-
-    numerical_transformer = Pipeline(steps=[
+    numerical_transformer = Pipeline([
         ('imputer', SimpleImputer(strategy='mean')),
         ('scaler', MinMaxScaler())
     ])
 
-    categorical_transformer = Pipeline(steps=[
+    categorical_transformer = Pipeline([
         ('imputer', SimpleImputer(strategy='constant', fill_value='missing')),
         ('onehot', OneHotEncoder(handle_unknown='ignore'))
     ])
 
-    preprocessor = ColumnTransformer(
-        transformers=[
-            ('num', numerical_transformer, numerical_cols),
-            ('cat', categorical_transformer, categorical_cols)
-        ])
+    preprocessor = ColumnTransformer([
+        ('num', numerical_transformer, numerical_cols),
+        ('cat', categorical_transformer, categorical_cols)
+    ], remainder='passthrough')
 
-    return preprocessor
-"""
-
-# build the full preprocessing and modeling pipeline
-def build_pipeline(df, drop_cols):
-    pipeline = ImbPipeline(steps=[
-        ('drop_columns', DropColumns(columns=drop_cols)),
-        ('preprocessor', preprocess_features(df)),
+    # Setup the full pipeline
+    pipeline = ImbPipeline([
+        ('preprocessor', preprocessor),
+        ('to_df', DataFrameEnsurer(preprocessor)), 
         ('resampler', RandomOverSampler(random_state=0)),
-        ('classifier', LogisticRegression(random_state=42))
+        ('classifier', LogisticRegression(random_state=42, max_iter=1000))  
     ])
-    return pipeline
 
+    return df, pipeline
+
+
+def perform_fairness_analysis(filepath):
+    df, pipeline = preprocess_data(filepath)
+    X_train, X_test, y_train, y_test = train_test_split(df.drop('y', axis=1), df['y'], test_size=0.2, random_state=42)
+    pipeline.fit(X_train, y_train)
 
-filepath = '/Users/sheidamajidi/Desktop/Winter2024/Winter2024-2/INSY695-076/Project/bank-full.csv'
-df = load_data(filepath)
-X_train, X_test, y_train, y_test = train_test_split(df.drop('y', axis=1), df['y'], test_size=0.2, random_state=42)
+    model = pipeline.named_steps['classifier']
+    preprocessed_data = pipeline.named_steps['to_df'].transform(pipeline.named_steps['preprocessor'].transform(X_train))
 
-# Build and train the pipeline
-drop_cols = [] # no columns to drop but for future if we want to drop columns
-pipeline = build_pipeline(df, drop_cols)
-pipeline.fit(X_train, y_train)
+    # predict_proba
+    predictions = model.predict_proba(preprocessed_data)[:, 1]  # probabilities for the positive class
 
+    # Audit the model using the probabilities of the positive class
+    total, _ = audit_model(lambda x: model.predict_proba(x)[:, 1], preprocessed_data) 
 
 
-# Fairness analysis
-# using the trained model for predictions and pass the whole test set for audit
-model = pipeline.named_steps['classifier']
-preprocessed_data = pipeline.named_steps['preprocessor'].transform(X_train)
+    fig = plot_dependencies(
+        total.median(),
+        title="FairML feature dependence",
+        fig_size=(10, 8)
+    )
+    plt.savefig('fairness_analysis.png')  
 
-# Audit the model
-total, _ = audit_model(model.predict_proba, preprocessed_data)
 
-# Plotting the results
-fig = plot_dependencies(
-    total.median(),
-    title="FairML feature dependence",
-    fig_size=(10, 8)
-)
-plt.show()
+if __name__ == '__main__':
+    perform_fairness_analysis('/Users/sheidamajidi/Desktop/Winter2024/Winter2024-2/INSY695-076/Project/bank-full.csv')