Employee Attrition Predictor

Overview

A comprehensive machine learning system that predicts employee attrition (voluntary turnover) using historical HR data. The project demonstrates full pipeline thinking — from exploratory data analysis and feature engineering through model selection, evaluation, and production-ready architecture design.

Employee attrition is a costly problem for organizations. Each departure triggers recruitment costs, onboarding time, and lost institutional knowledge. By predicting which employees are at risk of leaving, HR teams can proactively address retention before it becomes a problem.

Tech Stack

Pipeline Architecture

The system follows a structured ML pipeline with discrete, reusable stages:

Feature Engineering

The raw HR dataset is transformed through a comprehensive feature engineering pipeline to extract maximum predictive signal:

• Categorical Encoding — Ordinal and one-hot encoding for variables like department, job role, education field, and business travel frequency.
• Numerical Scaling — StandardScaler normalization for features with different scales (salary, age, distance from home, years at company).
• Feature Selection — Correlation analysis and mutual information scores to identify the most predictive features and remove multicollinear variables.
• Class Balancing — SMOTE (Synthetic Minority Over-sampling Technique) to address the inherent class imbalance — attrition is typically a minority event (~16% of employees).
• Derived Features — New features computed from existing ones: satisfaction ratios, tenure-normalized income, promotion gap indicators.

Model Selection

Multiple classification algorithms were evaluated using stratified k-fold cross-validation to find the best performer for this specific dataset:

• Logistic Regression — Baseline linear model with L2 regularization. Provides interpretable coefficients for understanding feature importance.
• Random Forest — Ensemble of decision trees with bagging. Strong performance on tabular data with built-in feature importance ranking.
• Gradient Boosting (XGBoost) — Sequential ensemble that builds trees to correct previous errors. Typically achieves highest accuracy on structured data.
• Support Vector Machine — RBF kernel SVM for capturing non-linear decision boundaries in the feature space.
• K-Nearest Neighbors — Instance-based learning used as an additional baseline for comparison.

Evaluation Strategy

Given the class imbalance (attrition is a minority class), accuracy alone is insufficient. The evaluation framework includes:

• Precision & Recall — Particularly focused on recall for the attrition class, since missing an at-risk employee (false negative) is more costly than a false alarm.
• F1 Score — Harmonic mean of precision and recall for balanced evaluation.
• ROC-AUC — Area under the ROC curve for threshold-independent model comparison.
• Confusion Matrix Analysis — Detailed breakdown of prediction types across both classes.
• Stratified K-Fold Cross-Validation — 5-fold CV maintaining class distribution in each fold for robust performance estimates.

Key Features

• Complete end-to-end ML pipeline from raw data to production-ready predictions
• 35+ engineered features including derived ratios and interaction terms
• SMOTE oversampling to handle class imbalance in attrition prediction
• Multi-model evaluation with stratified cross-validation
• Feature importance analysis identifying key attrition risk factors
• API-ready architecture with serialized model artifacts for deployment