A Comparative Machine Learning Approach For Predicting Employee Attrition In Human Resource Analytics

Abstract

Employee attrition imposes substantial recruitment, onboarding, and knowledge-transfer costs on organizations, making early identification of at-risk employees a strategic priority for human resource management. Prior work on this benchmark commonly reports a single best-performing classifier, yet rarely compares how competing algorithms trade precision against recall within one reproducible pipeline. This study compares the predictive performance of three supervised machine learning algorithms, namely Logistic Regression, Random Forest, and Support Vector Machine (SVM), for predicting employee attrition using the IBM HR Analytics Employee Attrition dataset, comprising 1,470 records and 31 predictor features. Data were encoded, standardized, and split 80:20 by stratified sampling preserving the 16.12% attrition ratio. Accuracy, precision, recall, F1-score, and AUC were reported, with emphasis on recall because an undetected leaver is costlier than a false alarm. Results show that Support Vector Machine achieved the highest recall (0.702) and F1-score (0.524), Logistic Regression achieved the highest accuracy (0.878) and AUC (0.805), while Random Forest, despite class-weight balancing, achieved the lowest recall (0.149) among the three models. Feature-importance analysis revealed that monthly income, age, total working years, years at company, and overtime status were the strongest predictive correlates of attrition, which signal associative rather than causal relevance. The contribution is a metric-level diagnosis showing that no algorithm dominates and that model suitability depends on the evaluation objective, so selection should follow organizational priorities rather than accuracy alone. As the data are a synthetic benchmark analysed on one split, the recall advantage of SVM is a benchmark-level indication requiring external validation.