Employee Attrition Prediction using Machine Learning

Goal: Predicting employee turnover and understanding the key drivers behind it using SHAP explainability

Business Impact

• Helps HR teams identify at-risk employees early

• Supports retention planning:

  • Reduce overtime for specific roles
  • Invest in improving satisfaction and engagement
  • Focus retention efforts on younger employees

• Potential to simulate attrition costs and optimize workforce strategy

About the Dataset

This project uses a fictional dataset created by IBM data scientists to simulate real-world HR data. The purpose is to uncover what factors contribute to employee attrition (i.e., leaving the company), and to help businesses proactively reduce turnover.

Source: IBM HR Analytics Attrition Dataset (Kaggle) Rows: 1470 Target variable: Attrition (Yes/No)

Key Features:

• Demographics: Age, Gender, MaritalStatus, Education
• Job Info: Department, JobRole, YearsAtCompany, JobSatisfaction
• Performance: PerformanceRating, OverTime, MonthlyIncome
• Label: Attrition — whether the employee left or stayed

Dashboard to visualize some features

dynamic dashboard can be found here for more interactivity and info in the tooltip

Exploratory Data Analysis

• Checked class imbalance: ~16% of employees left (Attrition=Yes)

• Explored patterns by marital status, satisfaction levels, income, and overtime

• Removed irrelevant or duplicate columns (e.g., EmployeeCount, Over18)

• Identified clear correlations between attrition and:

  • Overtime work
  • Younger age
  • Low satisfaction
  • High distance from home

Preprocessing & Feature Engineering

• Label Mapping: Attrition, Gender, OverTime, and Over18 were mapped to binary values
• Ordinal Encoding: BusinessTravel
• One-Hot Encoding: Applied to Department, EducationField, JobRole, MaritalStatus
• Scaling: StandardScaler applied to continuous numerical features
• Balancing: Used RandomOverSampler to handle class imbalance in training data

Modeling

Trained and evaluated six classification models:

Model	Accuracy	Key Observations
Logistic Regression	76%	Good recall on attrition class (66%)
K-Nearest Neighbors	67%	Weak precision/recall for attrition
Decision Tree	74%	Low performance on class 1
Random Forest	85%	High accuracy, weak recall for attrition
Support Vector Machine	60%	High recall but very low precision
XGBoost	86%	Best overall performance, selected for SHAP

Model Explainability with SHAP

Used SHAP (SHapley Additive exPlanations) to understand why the model predicts attrition.

Key Drivers of Attrition:

Feature	Insight
OverTime	Strongest predictor of attrition — frequent overtime increases risk
Age	Younger employees more likely to leave
EnvironmentSatisfaction	Dissatisfied employees more likely to leave
DistanceFromHome	Longer commutes correlate with higher attrition
JobSatisfaction	Low job satisfaction strongly tied to attrition

SHAP Summary Plot:

• Shows direction and strength of impact for each feature
• Confirms that the model aligns with real-world HR logic

🔝 Top Positive Drivers (Increase Attrition Risk)	🔽 Top Negative Drivers (Reduce Attrition Risk)
OverTime – Working overtime increases risk.	Older Age – Older employees are more stable.
Low Job Satisfaction – Unhappy in role.	Higher Monthly Income – More financial comfort.
Low Environment Satisfaction – Poor work environment.	Higher Stock Option Level – Incentivized to stay.
More Companies Worked For – Suggests instability.	Higher Job Involvement – More engaged employees.
Low Relationship Satisfaction – Poor manager/peer relationships.	Longer Years With Current Manager – Manager stability helps.

SHAP Dependence Plot:

• Interaction between DistanceFromHome and OverTime:

  • Employees living close by are often assigned overtime
  • Those far away + overtime = higher attrition

---

Web App: Employee Attrition Predictor (for HR decision making)

To make the solution actionable for HR teams, I developed a Streamlit web app that predicts attrition risk for individual employees based on input features like age, job satisfaction, overtime, income, and more.

Features of the App:

• Dynamic form for entering employee details
• Real-time prediction of attrition risk (Yes/No)
• Probability score showing model confidence
• Built-in explanations of the key risk and retention factors (based on SHAP findings)

ML Backend:

• Model: XGBoost (best performance, 86% accuracy)
• Preprocessing: One-hot encoding, scaling with StandardScaler
• Class balancing handled with RandomOverSampler

Why It Matters:

• Empowers HR teams to proactively assess and mitigate employee churn
• Encourages data-driven retention strategies
• Bridges the gap between machine learning output and business decision-making

Screenshot: you can find the interactive tool here

▶ To Run Locally:

pip install streamlit pandas numpy scikit-learn joblib xgboost
streamlit run Employee_Attrition_App.py