🧠 AttritionIQ — Employee Attrition Predictor

Predict whether an employee is likely to leave — with AI-powered explanations using SHAP.

---

📌 Overview

AttritionIQ is a machine learning web app that predicts employee attrition risk using the IBM HR Analytics dataset. It not only predicts whether an employee is likely to quit, but also explains why using SHAP (SHapley Additive exPlanations) — making it interpretable for HR decision-making.

---

🚀 Live Demo

👉 attritioniq-yashdivya5.streamlit.app

---

✨ Features

• Single Prediction — Enter employee details via an interactive form and get instant attrition risk with probability score
• Batch Prediction — Upload a CSV of employees and get predictions for all of them at once, with a downloadable results file
• SHAP Explanations — Waterfall plot showing which features drove each individual prediction
• Interpretable ML — Every prediction comes with a reason, not just a number

---

🛠️ Tech Stack

Layer	Technology
Model	XGBoost (v1.7.6)
Explainability	SHAP (v0.42.1)
Frontend	Streamlit
Data Processing	Pandas, Scikit-learn
Dataset	IBM HR Analytics (Kaggle)
Deployment	Streamlit Cloud

---

📊 Model Performance

Metric	Score
Accuracy	81.3%
F1 Score (Attrition class)	0.476
ROC-AUC	0.802
Recall (Attrition class)	53%

Note: Class imbalance handled using scale_pos_weight and a tuned prediction threshold of 0.35 to maximize recall on the minority (attrition) class.

---

📁 Project Structure

AttritionIQ/
├── streamlit_app.py       # Main Streamlit app
├── src/
│   └── preprocess.py      # Data loading & preprocessing
├── data/
│   └── WA_Fn-UseC_-HR-Employee-Attrition.csv
├── model/                 # Auto-generated at runtime
│   ├── xgb_model.pkl
│   ├── shap_summary.png
│   └── shap_waterfall.png
├── requirements.txt
└── README.md

---

⚙️ Run Locally

# Clone the repo
git clone https://github.com/yashdivya5/AttritionIQ.git
cd AttritionIQ

# Create virtual environment
python -m venv venv
venv\Scripts\activate  # Windows

# Install dependencies
pip install -r requirements.txt

# Run the app
streamlit run streamlit_app.py

---

🔍 How It Works

1. Preprocessing — Categorical encoding, dropping irrelevant columns, train/test split with stratification
2. Training — XGBoost with class imbalance handling (scale_pos_weight = 5.19)
3. Threshold Tuning — Default 0.5 threshold lowered to 0.35 to improve recall on attrition cases
4. Explainability — SHAP TreeExplainer generates feature-level attributions for every prediction

---

📈 Key Insights from SHAP

Top features driving attrition predictions:

• OverTime — strongest predictor of attrition
• MonthlyIncome — lower income → higher risk
• YearsAtCompany — early tenure employees leave more
• JobSatisfaction — lower satisfaction = higher attrition
• StockOptionLevel — no stock options = higher risk

---

👤 Author

Yash — Final year engineering student at NIE Mysore
GitHub • LinkedIn