Fairness-constrained explainable machine learning with differentially private federated training for employee attrition prediction in civil engineering and law

Abstract

Employee attrition in knowledge-intensive sectors such as civil engineering and law presents compounding risks to project continuity, regulatory compliance, and organisational profitability. Despite growing adoption of machine learning (ML) in human resource (HR) analytics, no prior study has simultaneously targeted these two professionally regulated sectors, incorporated employment-law fairness requirements (Title VII, Colorado AI Act 2024, California Civil Rights Council Regulations 2025, and the EU AI Act) as explicit optimisation constraints, or embedded federated learning to safeguard sensitive HR data. This study proposes FAIR-SEAP (Fairness-Aware Integrated Retention framework for Sector-stratified Engineering and Legal Attrition Prediction), an end-to-end framework that couples advanced ML with multi-metric explainability. The framework is evaluated on eight datasets: four publicly available cross-industry benchmarks, two sector-specific datasets constructed from published aggregate workforce statistics and expanded using a Conditional Tabular Generative Adversarial Network (CTGAN), and two federated partitions derived from them. It deploys Bayesian hyperparameter optimisation via the Tree-structured Parzen Estimator (TPE) with 3,000 trials, applies Demographic Parity (DP) and Equalised Odds (EO) as hard fairness constraints, and integrates SHAP, LIME, and Counterfactual Explanations (CFE) for global, local, and contrastive interpretability. On the civil engineering cohort, the Stacked Gradient Boosting–Histogram Gradient Boosting (SGB-HGB) ensemble attains 93.47% overall accuracy, an attrition-class F1-score of 82.94% (95% CI 82.10–83.78) and an AUC-ROC of 0.971; on the legal cohort the Transformer-Augmented AdaBoost (TA-AB) model attains 92.15% accuracy, an attrition-class F1-score of 81.75% (95% CI 80.83–82.67) and an AUC-ROC of 0.963. Both models exceed every baseline under paired two-sided t-testing with Holm–Bonferroni correction, with all adjusted p < 0.001; on the legal cohort TA-AB further exceeds SGB-HGB, the next-best model, by 1.40 F1 points at an adjusted p of 0.0011. SHAP analysis identifies Site Hazard Exposure, Certification Level and Project Phase Intensity as the leading attrition drivers in the civil engineering cohort, and Billable Hours Deviation, Mentor Score and perceived-equity (DEI) Score as the leading drivers in the legal cohort. These attributions describe statistical association rather than causal effect, and the reported fairness margins indicate conformity with pre-specified statistical thresholds rather than legal compliance, which requires contextual assessment beyond the scope of this study. Because neither sector cohort contains observed employee records and approximately 76% of each is CTGAN-generated, these figures characterise performance on simulated cohorts calibrated to published statistics rather than externally validated workforce performance. The study contributes a reproducible, sector-stratified and fairness-constrained analytical pipeline for HR practitioners in regulated professional sectors.