Intelligent Workforce Analytics: Predicting Employee Attrition Through Machine Learning

Employee attrition represents one of the most consequential and underestimated costs in modern technology organizations. While most HR departments track headcount and exit interview data, they lack the predictive infrastructure to identify at-risk employees before their resignation decisions become irreversible. This study addresses this gap by developing a production-grade machine learning pipeline for attrition risk prediction, which is applied to a workforce dataset modeled on Palo Alto Networks.

Using a dataset of 1,470 employee records containing demographic, compensation, satisfaction, and engagement attributes, we developed a comprehensive ML pipeline incorporating eight engineered behavioral features, multi-model evaluation across six classifiers, and a systematic threshold-optimization protocol. Our champion model, Logistic Regression with an optimized decision threshold of 0.30, achieved a ROC-AUC of 0.7844, an attrition recall of 72.34%, a Macro F1 of 0.5995, and a Mean CV F1 of 0.8500 on a held-out test set of 294 employees.

A key finding of this study is that linear decision boundaries, when combined with rigorous behavioral feature engineering and SMOTETomek resampling, outperform complex ensemble methods on small, imbalanced HR datasets. This counterintuitive result has direct implications for how organizations should approach predictive HR analytics in data-scarce conditions. The pipeline delivers a three-tier risk scoring framework (High / Medium / Low Risk) and a deployable Streamlit dashboard, enabling department-level attrition monitoring and what-if scenario exploration.

Keywords: Employee Attrition Prediction, Logistic Regression, Feature Engineering, SMOTETomek, HR Analytics, Class Imbalance, Threshold Optimization, ROC-AUC, Workforce Risk Scoring, Palo Alto Networks