FairHR-X: An Explainable and Bias-Aware Artificial Intelligence Framework for Transparent Employee Recruitment and Selection

Abstract

Artificial intelligence can improve recruitment efficiency, but models trained on historical hiring data may reproduce direct and proxy discrimination. This paper presents FairHR-X, an explainable and bias-aware lifecycle framework for human-supervised employee recruitment and selection that integrates predictive modelling, fairness auditing, mitigation, calibration, explainability, human oversight, and continuous monitoring. The framework was evaluated on 10,000 recruiting records regenerated from the CDEI data-generating process, using 6,000 training, 2,000 validation, and 2,000 held-out test observations. Logistic regression achieved the best predictive performance, with a test ROC-AUC of 0.941, PR-AUC of 0.911, and accuracy of 87.35%. The selected fairness-aware configuration, based on intersectional reweighing and Platt calibration, retained 86.45% accuracy and reduced race demographic-parity difference from 0.263 to 0.213, race equalised-odds difference from 0.090 to 0.042, sex demographic-parity difference from 0.163 to 0.119, and the intersectional demographic-parity gap from 0.436 to 0.340. Bootstrap analysis confirmed stable fairness improvements, while SHAP rankings showed high robustness across 1,000 resamples. For external validation, a recruiter-grouped protocol is defined for the 2026 PATHS2INCLUDE dataset comprising 15,036 candidate evaluations from 2,506 recruiters across four European countries. FairHR-X is intended as a transparent decision-support and governance framework rather than an autonomous hiring authority.