Data-Driven Workforce Analytics Using Supervised Learning Models in HRM

Abstract

Attrition by employees poses several challenges in workforce management, recruitment, knowledge management, and cost aspects of organizations. Workforce analytics can be used in HRM to understand trends from the collected employee data and determine the employment profile that is prone to attrition. In this paper, four supervised learning models—logistic regression, decision tree, random forest, and gradient boosting—are applied on the IBM HR analytics dataset on employee attrition and performance. The IBM dataset has 1,470 records and 35 variables with attrition being in the form of binary. Attrition among 237 employees (16.12%) was observed using exploratory analysis. More attrition cases were observed among employees working overtime compared to those not working overtime. Sales Representatives had the highest role attrition rate. Excluding the target, employee ID, and constant variables, 30 predictors were included. For evaluation, a stratified train-test split of 80% to 20%, as well as five-fold stratified cross-validation on the training set, was applied. The highest test F1-score (0.451) and recall (0.638) were recorded for Logistic Regression, while the highest accuracy (0.861) and ROC-AUC (0.805) were scored by Gradient Boosting. Overtime was the most important predictor using permutation importance, as identified for the chosen Logistic Regression model. It demonstrates the importance of evaluating the performance of the minority class in addition to accuracy. Since the data used in the analysis are synthetic, the results need to be interpreted as an exercise.