Artificial intelligence is increasingly used to support workforce planning, but employee attrition models raise difficult questions about privacy, fairness and whether predictions can be explained to the people affected by them.
A new peer-reviewed study proposes a machine-learning framework designed to address all three issues at once. In simulated cohorts representing civil engineering and legal-sector employees, the best-performing models achieved overall prediction accuracy above 92%, while incorporating explicit statistical fairness constraints, privacy-preserving federated training and multiple forms of explainability.
The results are technically strong, but they require an important qualification. The sector-specific cohorts did not contain observed employee records. Approximately 76% of each cohort was generated synthetically using a conditional generative model. The findings therefore show how the framework performed on simulated populations calibrated to published workforce statistics, not how accurately it would predict attrition in real engineering firms or law practices.
A framework built around fairness and privacy
Researchers Rashmi Kumari and colleagues developed a framework called FAIR-SEAP, short for Fairness-Aware Integrated Retention framework for Sector-stratified Engineering and Legal Attrition Prediction.
The study was published in Scientific Reports on 25 September 2026. It focuses on employee attrition in civil engineering and law, two professional settings where staff turnover can affect project continuity, client relationships, regulatory responsibilities and organisational costs.
Rather than optimising prediction accuracy alone, the researchers incorporated demographic parity and equalised odds as explicit fairness constraints. These statistical measures are intended to identify or limit differences in model outcomes across demographic groups. The authors also designed the framework with employment-law and responsible-AI considerations in mind, including regulatory developments in the United States and European Union.
However, satisfying a statistical fairness threshold is not the same as complying with employment law. The researchers explicitly caution that legal compliance depends on context and cannot be established by model metrics alone.
How the study was conducted
The framework was evaluated using eight datasets. Four were publicly available cross-industry benchmark datasets. Two sector-specific datasets represented civil engineering and legal workforces, while another two datasets were federated partitions derived from those sector cohorts.
The sector-specific data were constructed from published aggregate workforce statistics and expanded using a Conditional Tabular Generative Adversarial Network, or CTGAN. This technique generates synthetic tabular records that statistically resemble the source information without representing actual individual employees.
The researchers evaluated several machine-learning approaches and conducted Bayesian hyperparameter optimisation using 3,000 trials. They also integrated three explainability methods: SHAP for broader feature attribution, LIME for local explanations and counterfactual explanations to show how changes in inputs could alter individual predictions.
Federated learning was used as part of the privacy design. Instead of requiring all data to be pooled centrally, federated approaches allow models to be trained across separate data partitions while limiting the movement of raw records. Differential privacy was also incorporated to reduce the risk that information about individuals could be inferred from the trained model.
Civil engineering model reached 93.47% accuracy
For the simulated civil engineering cohort, a stacked Gradient Boosting and Histogram Gradient Boosting ensemble produced the strongest results. It achieved overall accuracy of 93.47% and an area under the receiver operating characteristic curve of 0.971.
Its F1-score for the attrition class was 82.94%, with a 95% confidence interval from 82.10% to 83.78%. The F1-score combines precision and recall and is particularly useful when the outcome classes are not evenly distributed.
The explainability analysis identified site hazard exposure, certification level and project phase intensity as the leading variables associated with attrition predictions in the civil engineering cohort.
These variables should not be interpreted as proven causes of employees leaving. Feature attribution describes how strongly a variable contributes to a model’s predictions within the analysed data. It does not establish that changing that factor would change an employee’s decision to leave.
Legal-sector model reached 92.15% accuracy
In the simulated legal-sector cohort, the strongest model was a Transformer-Augmented AdaBoost approach. It achieved overall accuracy of 92.15%, an attrition-class F1-score of 81.75% and an AUC-ROC of 0.963.
The 95% confidence interval for the attrition F1-score ranged from 80.83% to 82.67%. The model outperformed the next-best approach by 1.40 F1 percentage points, with an adjusted probability value of 0.0011.
Across model comparisons, the researchers used paired two-sided statistical tests with Holm-Bonferroni correction for multiple comparisons. The leading models exceeded the evaluated baselines with adjusted probability values below 0.001.
For the legal cohort, the leading model-attributed predictors included deviation in billable hours, mentor score and a perceived-equity measure related to diversity, equity and inclusion.
Why fairness constraints matter in workplace AI
Employee analytics can create governance risks because predictions may influence retention programmes, managerial attention, promotion opportunities or other employment decisions. A model that performs well overall can still generate systematically different error rates or outcomes across demographic groups.
The study’s approach is notable because fairness is treated as part of the optimisation problem rather than as an audit performed only after a model has been trained. In principle, this makes trade-offs between predictive performance and group-level statistical fairness visible during model development.
The addition of explainability also addresses a separate governance problem. A highly accurate attrition score may be difficult to justify if managers cannot understand which variables contributed to it. Combining global, local and counterfactual explanations provides different views of model behaviour, although explanations of a model are not necessarily explanations of human behaviour.
The synthetic data limitation is substantial
The strongest limitation is that the civil engineering and legal cohorts were not collections of observed individual employee records. Approximately 76% of each sector cohort was CTGAN-generated. The authors therefore describe the reported performance as applying to simulated cohorts calibrated against published aggregate statistics.
This matters because synthetic data can reproduce statistical patterns that were built into the generation process without capturing the full complexity of real workplaces. Organisational culture, labour-market conditions, management practices, compensation structures and personal circumstances may interact in ways that aggregate statistics cannot represent.
The models consequently require external validation using real-world workforce data before their reported accuracy, fairness or practical value can be assumed to generalise to employers.
Fairness metrics also have inherent limits. Demographic parity and equalised odds quantify particular statistical relationships, but no single metric can determine whether an employment practice is substantively fair. Different fairness criteria can conflict, and legal requirements vary by jurisdiction and use case.
What the study adds
The study offers a reproducible example of how employee attrition modelling can be designed around more than predictive accuracy. It combines sector-specific modelling, fairness constraints, explainability, federated learning and differential privacy in one analytical pipeline.
For organisations considering AI in human-resource management, the broader implication is that governance requirements can be incorporated into model design from the beginning. Accuracy remains important, but privacy, interpretability and fairness can be treated as measurable design objectives rather than secondary considerations.
The next test is considerably harder. The framework needs evaluation on genuine workforce records from multiple organisations and jurisdictions, with prospective assessment of whether its predictions remain accurate, fair and useful under real employment conditions.
Source Information
Study: Fairness-constrained explainable machine learning with differentially private federated training for employee attrition prediction in civil engineering and law
Authors: Rashmi Kumari, Suresh Pratap, Pradyut Anand, Priyanka Anand, Mamuye Busier Yesuf and colleagues
Journal: Scientific Reports
Published: 25 September 2026
DOI: 10.1038/s41598-026-71406-4
Study design: Machine-learning framework development and evaluation using benchmark, synthetic sector-specific and federated datasets
