Optimizing Multi-Class Imbalance in Employee Attrition Prediction Using Non-dominated Sorting Genetic Algorithm II with Fairness-Driven Fitness Function

Abstract

Class imbalance in multi-class employee attrition prediction is a major challenge in human resource analytics, causing bias towards the majority class and poor performance on the minority class. This study proposes a multi-objective genetic algorithm (GA) framework using NSGA-II to address multi-class imbalance in the IBM HR Analytics Employee Attrition & Performance dataset. By optimising precision, recall, and fairness (normalised Demographic Parity Difference), this framework generates synthetic samples for minority classes (Resign, Retire_Termination) through two-class adaptive clustering and a weighted fitness function (β=0.3). Experiments were conducted with the XGBoost classifier, comparing GA with the SMOTE baseline. Results show that GA achieves a macro F1-score of 0.65 ± 0.02, surpassing SMOTE (0.56 ± 0.03), with significant improvements in Resign (F1-score 0.59 vs. 0.51) and Retire_Termination (F1-score 0.42 vs. 0.24). The fairness value of GA (0.82 ± 0.02) was higher than that of SMOTE (0.75 ± 0.03), indicating fairer predictions. Visualisation of the Pareto front and convergence of GA illustrates the trade-off between objectives and algorithm robustness. Key contributions include a GA framework that integrates fairness, advantages over SMOTE, and flexibility for HR applications.