Employee Attrition Prediction

Using the IBM HR Analytics Employee Attrition & Performance dataset, I built interpretable machine learning models to understand why employees leave and to help HR teams identify at-risk staff. This project combines EDA, class imbalance handling, and predictive modeling with a focus on interpretability.

Problem

Employee turnover is costly, and HR teams need tools to identify risk factors and predict attrition.
The question: Which factors most strongly influence attrition, and how can we build a model that balances accuracy with interpretability?

Approach

• EDA – Explored distributions of satisfaction, overtime, department, tenure, and job level with respect to attrition.
• Preprocessing –
  • One-hot encoding for categorical features
  • SMOTE to handle class imbalance
  • Binning (Age, YearsAtCompany)
  • Rare label encoding for JobRole
• Modeling –
  • Logistic Regression (baseline)
  • Random Forest
  • XGBoost
  • Hyperparameter tuning with RandomizedSearchCV
• Evaluation –
  • Confusion Matrix, ROC Curve, AUC
  • Classification Report (Precision, Recall, F1)

Results

• Best-performing model: XGBoost with SMOTE
• Most predictive features: OverTime, StockOptionLevel, JobSatisfaction, YearsAtCompany

Impact

• Addressed class imbalance using SMOTE, which significantly improved recall on the minority (attrition) class.
• Showed that behavioral features (e.g., job satisfaction, overtime, work-life balance) are critical for predicting attrition.
• Delivered an interpretable workflow that HR teams could apply to support retention strategies.

Skills & Tools

• Python (pandas, NumPy, scikit-learn, XGBoost, imbalanced-learn)
• EDA and feature engineering
• Handling class imbalance (SMOTE)
• Model evaluation (AUC, recall, precision, F1)
• Hyperparameter tuning (RandomizedSearchCV)
• Visualization (matplotlib, seaborn)

View the project on GitHub