Modelpublic

024vishnu/adult-income-responsible-ai

sourceHugging Facemitupdated 2mo agoView on Hugging Face

0likes

Model Card

Adult Income Classification and Fairness Model

Model Details

This educational project predicts Class-label = 1 in the cleaned Adult dataset and compares Random Forest, Multinomial Naive Bayes, k-nearest neighbors, and Linear SVM.

—Selected model: Random Forest
—Framework: scikit-learn 1.6.1
—Training timestamp: 2026-08-03T10:36:46.709070+00:00
—Sensitive attribute evaluated: gender

Intended Use

Responsible AI education, model-card demonstrations, and classroom experiments with predictive performance and group fairness.

Out-of-Scope Use

Do not use this model for employment, lending, insurance, housing, healthcare, criminal-justice, or other consequential decisions.

Dataset

—Rows: 45,222
—Input features: 14
—Target: Class-label
—Positive class: 1
—Protected groups evaluated: Male and Female

Preprocessing

—Stratified 80/20 train/test split
—Median imputation and min-max scaling for numeric columns
—Most-frequent imputation and one-hot encoding for categorical columns

Evaluation Results

Model	Accuracy	Male positive rate	Female positive rate	Statistical parity difference	Disparate impact (Female/Male)
Random Forest	0.8515	0.2707	0.0811	0.1896	0.2995
Linear SVM	0.8452	0.2541	0.0744	0.1797	0.2928
k-NN	0.8262	0.2803	0.0814	0.1989	0.2905
Naive Bayes	0.7925	0.4537	0.0640	0.3897	0.1411

Fairness Evaluation

Statistical parity difference: P(predicted positive | Male) - P(predicted positive | Female). A value near zero indicates similar positive-prediction rates. Disparate impact is reported as the female rate divided by the male rate.

Limitations and Risks

—Statistical parity is only one group-fairness metric and does not prove overall fairness.
—The dataset uses binary gender categories and does not represent all gender identities.
—Historical data may encode structural inequalities.
—Accuracy can hide unequal false-positive and false-negative rates.
—No causal, intersectional, robustness, calibration, or post-deployment analysis was performed.

Recommended Further Evaluation

Evaluate equal opportunity, equalized odds, calibration, subgroup confusion matrices, intersectional groups, robustness, stakeholder-defined harms, and post-deployment monitoring.

Reproducibility

The complete pipeline is stored in adult_income_best_model.joblib, and the comparison is stored in model_comparison.csv.

Environmental Impact

These are relatively small CPU-based tabular models. No formal energy or carbon measurement was conducted.