Fairness-Aware Machine Learning for Employee Attrition Prediction: An Empirical Study

Research Question: How do different machine learning models compare in predictive performance and algorithmic fairness for employee attrition prediction, and can bias mitigation reduce gender disparity without significant accuracy loss?

---

Research Contributions

This project makes four concrete research contributions:

1. Comparative ML evaluation — Four models (Logistic Regression, Random Forest, XGBoost, Neural Network) evaluated with Accuracy, F1, and AUC on an imbalanced HR dataset
2. Cross-model fairness auditing — Disparate Impact, Statistical Parity Difference, and Equal Opportunity Difference computed per model, revealing a measurable accuracy-fairness tradeoff
3. Bias mitigation experiment — Reweighing (Kamiran & Calders, 2012) applied to quantify fairness improvement vs accuracy cost
4. Explainability analysis — SHAP-based feature importance connecting model behaviour to fairness outcomes

---

Key Findings

Model	Accuracy	F1 Score	AUC	Disparate Impact	Verdict
Logistic Regression	0.839	0.000	0.636	—	Predicts no attrition (accuracy paradox)
Neural Network (MLP)	0.814	0.163	0.615	1.495	⚠️ Fairness risk
Random Forest	0.807	0.175	0.583	1.252	⚠️ Fairness risk
XGBoost	0.791	0.193	0.565	1.274	⚠️ Fairness risk

Finding 1 — Accuracy paradox: Logistic Regression scores highest accuracy (83.9%) but F1=0, meaning it never predicts attrition. A model can appear accurate by always predicting the majority class in a 5.2:1 imbalanced dataset.

Finding 2 — Accuracy-fairness tradeoff: Models that actually predict attrition all show Disparate Impact > 1.25 (over-predicting attrition for females relative to males). Correlation between AUC and fairness deviation = 0.90.

Finding 3 — Mitigation result: Reweighing reduced accuracy cost to only Δ=−0.009 but did not improve Disparate Impact, suggesting gender disparity is driven by feature-level correlations (income, job level) rather than representation imbalance — motivating future feature-level debiasing work.

---

System Architecture

flowchart TD
    A[IBM HR Dataset\n1470 employees] --> B[Data Preprocessing\nLabel Encoding · Train/Test Split\n70/30 · Stratified]

    B --> C[ML Training Pipeline\ntrain_model.py]
    C --> C1[Logistic Regression]
    C --> C2[Random Forest]
    C --> C3[XGBoost]
    C --> C4[Neural Network MLP]

    C1 & C2 & C3 & C4 --> D[Fairness Analysis\nfairness_analysis.py]
    D --> D1[Disparate Impact]
    D --> D2[Statistical Parity Diff]
    D --> D3[Equal Opportunity Diff]

    D --> E[Bias Mitigation\nfairness_mitigation.py]
    E --> E1[Reweighing\nKamiran & Calders 2012]
    E1 --> E2[Fair Random Forest]

    C2 --> F[Explainability\nSHAP LinearExplainer]
    F --> F1[Global Feature Importance]

    D & E & F --> G[Streamlit Dashboard\napp.py · 6 tabs]
    G --> G1[📊 Analytics]
    G --> G2[📈 Model Comparison]
    G --> G3[⚖️ Fairness Analysis]
    G --> G4[🔬 Fairness Mitigation]
    G --> G5[🧠 Explainability]
    G --> G6[🤖 Prediction]

Loading

---

Project Structure

responsible-ai-attrition-dashboard/
├── app.py                        # Streamlit dashboard (6 tabs)
├── train_model.py                # Multi-model training pipeline
├── fairness_analysis.py          # Per-model fairness evaluation
├── fairness_mitigation.py        # Reweighing bias mitigation
├── requirements.txt
├── data/
│   └── WA_Fn-UseC_-HR-Employee-Attrition.csv
├── models/                       # Generated — one .pkl per model
│   ├── logistic_regression.pkl
│   ├── random_forest.pkl
│   ├── xgboost.pkl
│   ├── neural_network_mlp.pkl
│   └── fair_random_forest.pkl
└── results/                      # Generated — CSVs + plots
    ├── model_comparison.csv
    ├── fairness_comparison.csv
    ├── fairness_comparison_plot.png
    ├── mitigation_comparison.csv
    └── mitigation_comparison_plot.png

---

Dataset

IBM HR Analytics Employee Attrition Dataset — 1,470 employees, 35 features.

Features used: Age · Gender · Education · Job Level · Monthly Income · Years at Company

Key characteristic: 5.2:1 class imbalance (No:Yes attrition) — directly impacts model behaviour and is central to the accuracy paradox finding.

Source: Kaggle — IBM HR Analytics

---

Fairness Metrics

Metric	Formula	Ideal	Threshold
Disparate Impact (DI)	P(ŷ=1|female) / P(ŷ=1|male)	1.0	0.8 – 1.25 (EEOC 80% rule)
Statistical Parity Diff (SPD)	P(ŷ=1|female) − P(ŷ=1|male)	0.0	|SPD| < 0.05
Equal Opportunity Diff (EOD)	TPR(female) − TPR(male)	0.0	|EOD| < 0.05

---

How to Run Locally

# 1. Clone
git clone https://github.com/asthasingh0660/responsible-ai-attrition-dashboard.git
cd responsible-ai-attrition-dashboard

# 2. Set up environment
python -m venv venv
venv\Scripts\activate        # Windows
# source venv/bin/activate   # Mac/Linux

# 3. Install dependencies
pip install -r requirements.txt
pip install xgboost

# 4. Run in order
python train_model.py           # trains all 4 models
python fairness_analysis.py     # fairness evaluation per model
python fairness_mitigation.py   # bias mitigation experiment

# 5. Launch dashboard
streamlit run app.py

---

Responsible AI Design

This system is explicitly designed as a decision-support tool, not an automated decision-maker.

• Human-in-the-loop — every prediction tab includes clear disclaimers
• Fairness-first evaluation — fairness metrics computed alongside performance metrics, not as an afterthought
• Honest mitigation reporting — mitigation results reported truthfully including cases where improvement was not achieved
• Explainability — SHAP values connect model behaviour to human-interpretable features
• Confidence-aware predictions — low-confidence outputs flagged explicitly

---

Technologies

Python · Scikit-learn · XGBoost · SHAP · Streamlit · Pandas · NumPy · Matplotlib · Seaborn

---

Author

Astha Singh

---

Disclaimer

This project is intended for educational and research purposes only. Predictions and insights should not be used as the sole basis for real-world HR decisions.