Evaluating the Effectiveness and Fairness of Machine Learning Models in Employee Recruitment and Selection Process

Abstract

The use of Artificial intelligence (AI) has been rapidly increased to support employee recruitment and selection process. The process employee recruitment and selection including candidate screening, ranking, assessment and hiring recommendations. The automated decision systems regarding employee recruitment and selection can improve consistency and scalability but they may also increase the inequalities associated with the candidate characteristics such as gender, age, location and others. The proposed paper presents a comparative machine-learning framework to evaluate both the predictive effectiveness as well as procedural fairness in automated employee selection. The simulation is carried out by using public AI Fair Recruitment Dataset (Kaggle, n.d.), which is consist of 121,170 candidate records after removal of 30 incomplete records and 16 variables. Five supervised machine learning models including Logistic Regression (LR), Decision Tree (DT), Random Forest (RF), a histogram-based gradient-boosting ensemble (XGBoost) and a multilayer-perceptron Artificial Neural Network (ANN) are used in the proposed comparative framework, These models are trained on a stratified 70/15/15 split and evaluated only test subset (n = 18,176). LR achieved perfect separation with accuracy = 100.0% and ROC-AUC = 1.000. From the performance of LR it is clear that the dataset's hiring label are generated by a near-deterministic linear rule over four scored attributes rather than by noisy, real-world hiring behavior. The permutation-based features confirmed that the Technical_Test_Score, Aptitude_Test_Score, Interview_Score, and Skill_Score together are effectively all predictive signal. Moreover, the Gender contributed negligibly and Age, Location, and Job_Role_Applied contributed is almost none. Subgroup auditing found modest, not severe, gender disparities across all five models (disparate impact ratio range from 0.880 to 0.972; demographic parity difference −0.007 to −0.033), all within the conventional four-fifths threshold. A feature-removal mitigation strategy that excluded Gender and Age from a random forest's predictor set narrowed the gender selection-rate gap from 2.59 to 0.05 percentage points (disparate impact ratio 0.896 to 1.002) at a small cost in F1-score (94.14% to 93.63%). The results illustrate that predictive effectiveness and demographic fairness are separate dimensions that must be audited jointly, and that even a synthetic, near-deterministic labeling process can carry small but measurable demographic disparities worth mitigating.