FairPath: Fair and Explainable Talent Promotion Prediction Using Counterfactual Explanations and Fairness Constraints
International Core Journal of Social Sciences · 29.10.2026 · 2 (2) · 19-26
DOI: 10.54097/bhp46k56 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Core Journal of Social Sciences, online erschienen am 29.10.2026. Verfasst von Elisa Sharma und 3 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY-NC 4.0).
Abstract (Originalsprache)
Machine-learning models that support promotion decisions are trained on historical HR records and can inherit the unequal treatment those records contain. Recent work such as FairPromote combines adversarial debiasing with SHAP attributions and evaluates outcome-level group fairness, but it does not tell a non-promoted employee what would change the decision, nor whether the effort required to obtain a promotion differs across groups. We present FairPath, which couples fairness constraints with counterfactual explanations. FairPath builds on the FairPromote adversarial encoder–predictor, replaces its loss-maximizing adversary with a confusion objective, and adds a counterfactual logit-pairing constraint that makes promotion scores invariant to flipping the sensitive attribute. For explanation we introduce TCAR, a model-agnostic generator of temporally consistent actionable recourse that searches a lattice of effort improvements and elapsed years exactly, advancing age, tenure and time in role jointly. TCAR also supports a counterfactual audit based on the attribute-flip rate (AFR) and the recourse-burden gap (RBG) between groups. On the IBM HR dataset and a 54,808-employee promotion dataset, under a controlled historical-label-bias protocol with clean test labels, FairPath reduced AFR from 10.2% (MLP) and 14.8% (FairPromote) to 0.1% on IBM-HR and to 0.0% on WNS, and lowered the IBM-HR recourse-burden gap of the MLP from 0.78 to 0.15 cost units without a significant loss of AUC. TCAR returned valid and feasible counterfactuals for every query, whereas none of the counterfactuals produced by DiCE or a gradient-based generator was both valid and feasible.
Zitieren
APA
Sharma, E., Hall, A., Price, H., & Moore, M. (2026). FairPath: Fair and Explainable Talent Promotion Prediction Using Counterfactual Explanations and Fairness Constraints. International Core Journal of Social Sciences, 2(2), 19-26. https://doi.org/10.54097/bhp46k56
Vancouver
Sharma E, Hall A, Price H, Moore M. FairPath: Fair and Explainable Talent Promotion Prediction Using Counterfactual Explanations and Fairness Constraints. International Core Journal of Social Sciences. 2026;2(2):19-26. doi:10.54097/bhp46k56
BibTeX
@article{Sharma2026,
author = {Sharma, Elisa and Hall, Avery and Price, Hiro and Moore, Miles},
title = {FairPath: Fair and Explainable Talent Promotion Prediction Using Counterfactual Explanations and Fairness Constraints},
journal = {International Core Journal of Social Sciences},
year = {2026},
volume = {2},
number = {2},
pages = {19--26},
doi = {10.54097/bhp46k56}
}
EUVIMED