Preparing Model Outputs for Fairness Analysis

Before conducting any fairness or bias audits with EquiBoots, we must first train a model and extract the necessary components: predicted labels, predicted probabilities, true labels, and sensitive attributes. This section walks through the process using the Adult Income dataset [1], a popular dataset from the UCI Machine Learning Repository [2].

Step 1: Install and Import Dependencies

We begin by installing the ucimlrepo package and importing the necessary Python libraries for data handling, preprocessing, modeling, and evaluation.

pip install ucimlrepo

## Import Necessary Libraries

from ucimlrepo import fetch_ucirepo
import pandas as pd
import numpy as np
from sklearn.preprocessing import LabelEncoder
from sklearn.model_selection import train_test_split
from xgboost import XGBClassifier
from sklearn.metrics import classification_report

Step 2: Load the Dataset

We fetch the Adult Income dataset directly from the UCI repository using ucimlrepo. This returns a dataset with features and a target variable indicating whether an individual’s income is above or below $50K.

## Fetch Dataset
adult = fetch_ucirepo(id=2)
adult = adult.data.features.join(adult.data.targets, how="inner")

Step 3: Clean the Data

Missing values are removed to simplify modeling. We also create a backup copy of the cleaned dataset.

a. Drop missing values

adult.dropna(inplace=True)

b. Copy DataFrame for posterity

df = adult.copy()

Step 4: Encode the Target Variable

The target column 'income' is a string. We convert it into a binary format: 0 for <=50K and 1 for >50K.

def outcome_merge(val):

    if val == '<=50K' or val == '<=50K.':
        return 0
    else:
        return 1

df['income'] = df['income'].apply(outcome_merge)

Step 5: Prepare Features and Labels

We split the dataset into features X and labels y. Categorical variables are encoded as pandas category types to be handled natively by XGBoost.

X = df.drop("income", axis=1)
y = df["income"]

for col in X.columns:
if isinstance(X[col], object):
    X[col] = X[col].astype("category")

Step 6: Train-Test Split

We split the data into training and test sets using an 80/20 ratio.

X_train, X_test, y_train, y_test = train_test_split(
    X,
    y,
    test_size=0.2,
    random_state=42,
)

Step 7: Train the XGBoost Model

We fit an XGBoost classifier on the training data.

model = XGBClassifier(
    eval_metric='logloss',
    random_state=42,
    enable_categorical=True
)
model.fit(X_train, y_train)

Note

Note that enable_categorical=True is used to support categorical columns directly.

Step 8: Generate Predictions and Evaluate

We obtain both predicted class labels (y_pred) and predicted probabilities (y_prob) from the model and print a standard classification report to evaluate performance.

y_pred = model.predict(X_test)
y_prob = model.predict_proba(X_test)
print(classification_report(y_test, y_pred))

               precision    recall  f1-score   support

           0        0.88      0.94      0.91      7170
           1        0.76      0.63      0.69      2355

    accuracy                            0.86      9525
   macro avg        0.82      0.78      0.80      9525
weighted avg        0.85      0.86      0.86      9525

Step 9: Extract Model Outputs

We now extract the model’s predicted class labels, predicted probabilities for the positive class (income > 50K), and convert the true labels to a NumPy array. These are the core inputs required by EquiBoots.

## Get predicted class labels (0 or 1)
y_pred = model.predict(X_test)

## Get predicted probabilities for class 1 (income > 50K)
y_prob = model.predict_proba(X_test)[:, 1]

## Convert ground truth labels to NumPy array
y_true = y_test.to_numpy()