FairLeave: Explainable Employee Attrition Prediction via Graph Neural Networks and Counterfactual Explanations
International Journal of Computational and Biological Sciences · 2026
DOI: 10.66238/ijcbs94 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Computational and Biological Sciences, 2026. Verfasst von Victor Lin. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Employee attrition prediction is a high-impact human resource analytics task, but conventional tabular models often treat employees as independent observations and provide limited guidance about how a high-risk prediction could be reduced through feasible organizational actions. This paper proposes FairLeave, a graph-based and explainable attrition prediction framework that represents employees as nodes in a similarity graph, learns neighborhood-aware representations with a residual graph convolutional network, audits gender-related disparity through an adversarial branch, and produces actionable counterfactual explanations for high-risk cases. The method is inspired by recent fairness-aware and explainability-aware HR analytics research, especially the FairPromote framework, which demonstrates the value of combining predictive modeling with auditable explanations in personnel decision support. Experiments are conducted on the IBM HR Analytics Employee Attrition & Performance dataset, a fictional but widely used benchmark containing 1,470 employee records. The sensitive attribute Gender is excluded from predictive inputs and used only for fairness auditing. Across three stratified splits, FairLeave achieves an AUC-ROC of 0.863±0.013 and AUC-PR of 0.633±0.007, outperforming the strongest non-graph baseline in AUC-ROC while reducing gender equalized odds difference compared with the multilayer perceptron baseline. Counterfactual evaluation on ten high-risk test employees reduces predicted attrition probability by 0.392 on average, with a median of three recommended actions and an 80% threshold-crossing success rate. These results show that graph structure and recourse-style explanations can make attrition analytics more transparent, auditable, and operationally useful, while the study remains bounded by the limitations of a fictional public benchmark
Zitieren
APA
Lin, V. (2026). FairLeave: Explainable Employee Attrition Prediction via Graph Neural Networks and Counterfactual Explanations. International Journal of Computational and Biological Sciences. https://doi.org/10.66238/ijcbs94
Vancouver
Lin V. FairLeave: Explainable Employee Attrition Prediction via Graph Neural Networks and Counterfactual Explanations. International Journal of Computational and Biological Sciences. 2026. doi:10.66238/ijcbs94
BibTeX
@article{Lin2026,
author = {Lin, Victor},
title = {FairLeave: Explainable Employee Attrition Prediction via Graph Neural Networks and Counterfactual Explanations},
journal = {International Journal of Computational and Biological Sciences},
year = {2026},
doi = {10.66238/ijcbs94}
}
EUVIMED