Explainable-AI applied to HR records finds overtime, travel and promotion gaps drive resignations more than pay; simulation suggests eliminating overtime could cut predicted attrition by about 17% and retain roughly 31 staff in the analyzed dataset, outperforming modest salary hikes.
Explainable AI for Employee Retention in Green Human Resource Management: Integrating Prediction, Interpretation, and Policy Simulation
An interpretable, calibrated logistic-regression model on IBM HR records shows non-monetary factors—particularly overtime, frequent travel, and limited promotion opportunities—predict attrition more strongly than salary, and counterfactual simulations suggest eliminating overtime could reduce predicted attrition by 17.35%, retaining roughly 31 employees in the sample.
Retaining the green workforce, employees driving sustainability and environmental innovation, is essential for organizational resilience and long-term environmental goals. While prior Green HRM research has primarily relied on survey-based methodologies and theoretical frameworks to examine retention factors, these approaches lack predictive capability and fail to provide actionable, employee-specific insights. This study advances beyond descriptive and correlational analyses by employing explainable artificial intelligence (XAI) to develop a transparent, data-driven framework for identifying attrition drivers and quantitatively evaluating retention strategies. Unlike existing studies that rely on self-reported perceptions, our approach leverages objective HR data and machine learning to predict individual-level attrition risk with calibrated probabilities. Leveraging the IBM HR Analytics dataset as a proxy for sustainability-focused roles, we construct an interpretable logistic regression model with strong predictive performance and isotonic regression calibration. Global and local interpretability techniques, including SHAP, LIME, and permutation importance, show that non-monetary factors, such as excessive overtime, frequent business travel, and limited promotion opportunities, have a greater impact on turnover risk than salary levels. These findings align with Green Human Management (Green HRM) principles, which emphasize work–life balance and employee well-being. Crucially, our policy simulation framework, absent from prior Green HRM studies, demonstrates that eliminating overtime could reduce predicted attrition probability by 17.35% for affected employees, potentially retaining 31 staff members, substantially outperforming modest salary adjustments. This work expands the value of predictive AI into HR analytics by consolidating HR analytics with Green HRM through a novel methodology that bridges the gap between prediction and actionable intervention. It represents the first systematic integration of XAI-based predictive modeling with counterfactual policy simulation in environmentally conscious sustainable organizations.
Summary
Main Finding
Using explainable AI on objective HR data, the study shows that non-monetary workplace factors (excessive overtime, frequent business travel, limited promotion opportunities) are stronger predictors of individual attrition risk than salary. A calibrated, interpretable logistic-regression model plus counterfactual policy simulations indicate that eliminating overtime could lower predicted attrition probability by 17.35% for affected employees and, in the dataset used, could potentially retain about 31 staff — a larger effect than modest salary increases. The approach moves retention analysis from descriptive correlations to actionable, employee-level predictions and policy evaluation.
Key Points
- Problem addressed: Prior Green HRM research relies on surveys and theory, lacking predictive power and actionable, employee-specific interventions.
- Data source: IBM HR Analytics dataset used as a proxy for sustainability-focused roles.
- Modeling approach: Interpretable logistic regression producing calibrated, individual-level attrition probabilities (isotonic regression used for calibration).
- Explainability toolkit: Global and local interpretability via SHAP, LIME, and permutation importance.
- Main drivers of attrition (higher influence than salary): overtime, business travel frequency, promotion opportunities.
- Policy simulation: Counterfactual experiments show eliminating overtime reduces predicted attrition by 17.35% for affected employees and could retain ~31 employees in the sample — outperforming modest pay increases.
- Novelty: First systematic integration of XAI-based predictive modeling with counterfactual policy simulation aimed at sustainability-oriented HR (Green HRM).
Data & Methods
- Dataset: IBM HR Analytics (treated as a proxy for green/sustainability roles; objective HR records rather than self-reports).
- Predictive model: Logistic regression selected for interpretability and transparency.
- Calibration: Isotonic regression applied to convert model scores to well-calibrated probabilities.
- Interpretability/explainability:
- Global importance: permutation importance, aggregated SHAP values.
- Local explanations: SHAP and LIME to explain individual predictions and to identify employee-specific intervention levers.
- Counterfactual policy simulation: Systematically modify features (e.g., set overtime to zero) and compute the change in each employee’s calibrated attrition probability to estimate impact and aggregate retained headcount.
- Validation: Model demonstrated strong predictive performance and well-calibrated probabilities (study reports model readiness for policy simulation; exact performance metrics not reproduced here).
Implications for AI Economics
- Economic value of targeted retention:
- More cost-effective allocation of retention resources by prioritizing non-monetary interventions with larger marginal impact (e.g., reducing overtime, limiting travel, improving promotion pathways).
- Potential reductions in recruiting, onboarding, and productivity losses associated with turnover — implying positive ROI for AI-driven HR interventions.
- Human-capital and productivity effects:
- Retaining "green" workforce preserves organizational capabilities critical for environmental innovation and long-term sustainability goals, generating positive firm-level externalities.
- Policy design and micro-targeting:
- Calibrated, employee-level predictions enable marginal-cost analyses and prioritization (which employees to target, which interventions to apply), improving intervention efficiency versus uniform, across-the-board policies.
- Market and labor implications:
- Widespread adoption could shift firm incentives toward non-wage retention investments, altering compensation bundles and labor-market competition for sustainability-focused talent.
- Methodological and implementation caveats:
- Counterfactual simulations are predictive, not inherently causal; estimated effects require causal validation (e.g., randomized trials) before large-scale policy rollout.
- Generalizability limits: IBM dataset used as proxy for sustainability roles — results may differ in true green-workforce populations or across industries/countries.
- Risks to consider: fairness and bias in models, privacy concerns with employee-level predictions, potential morale effects if interventions are unevenly applied.
- Research and policy agenda:
- Next steps include causal evaluation of recommended interventions, cost–benefit analyses that include implementation costs, and replication in actual sustainability-role datasets.
- Regulators and firms should combine XAI-driven predictions with ethical governance and randomized pilots to translate predictive gains into verified economic and environmental outcomes.
Assessment
Paper Typecorrelational
Evidence Strengthlow — Findings rest on well-specified predictive models and calibration on objective HR data, which supports strong predictive claims, but the counterfactual policy simulations are not supported by causal identification (no randomization, natural experiment, or instrumental strategy), rely on a single proprietary dataset used as a proxy for green/sustainability roles, and lack external or experimental validation—so claims about the effect of interventions on actual turnover are speculative.
Methods Rigormedium — The study applies appropriate and transparent methods for prediction and interpretability (logistic regression for transparency, isotonic calibration, SHAP/LIME, permutation importance) and performs individual-level counterfactual simulations, which is methodologically sound for predictive analysis; however, reporting omits key causal identification, the summary does not provide full performance metrics or robustness checks, the dataset is a proxy sample with potential selection/context biases, and no causal validation (RCTs or quasi-experimental checks) is presented.
SampleEmployee-level records from the IBM HR Analytics dataset (used as a proxy for sustainability/green roles), containing HR features such as overtime hours, business-travel frequency, promotion history, salary and other demographics/workplace variables; exact sample size and time frame not reported in the summary.
Themeslabor_markets productivity
IdentificationNo causal identification: uses an interpretable, calibrated logistic-regression model on observational IBM HR records and conducts counterfactual feature manipulations (e.g., set overtime to zero) to estimate predictive changes in attrition probability; authors explicitly note these are predictive/simulation results rather than causal estimates.
GeneralizabilitySingle-source dataset (IBM) used as a proxy for sustainability-focused roles — may not represent other firms, sectors, or true 'green' workforces, Organizational context and HR practices at IBM may differ from smaller firms or different countries/industries, Observational, cross-sectional (or non-experimental) data limit transferability of estimated policy effects to real-world intervention outcomes, Unclear sample size/time period and potential demographic/selection biases reduce external validity, Cultural, legal, and labor-market differences (e.g., overtime norms, labor protections) constrain applicability across jurisdictions
Claims (15)
| Claim | Direction | Confidence | Outcome | Details |
|---|---|---|---|---|
| Non-monetary workplace factors (excessive overtime, frequent business travel, limited promotion opportunities) are stronger predictors of individual attrition risk than salary. Turnover | positive | medium | individual attrition risk (predicted probability of attrition) |
non-monetary factors (overtime, travel, limited promotion) ranked higher than salary via SHAP/permutation importance
0.09
|
| An interpretable logistic-regression model, calibrated with isotonic regression, produces well-calibrated, individual-level attrition probabilities suitable for policy simulation. Turnover | null_result | medium | calibrated predicted probability of attrition (model calibration/readiness) |
logistic regression calibrated with isotonic regression produces well-calibrated individual attrition probabilities
0.09
|
| Eliminating overtime could lower predicted attrition probability by 17.35% for affected employees (per the model's counterfactual simulation). Turnover | negative | medium | change in calibrated predicted attrition probability (percentage point reduction) |
counterfactual: eliminating overtime lowers predicted attrition probability by 17.35% (average)
0.09
|
| In the dataset used, eliminating overtime could potentially retain about 31 employees — a larger effect than modest salary increases. Turnover | negative | medium | predicted retained headcount (number of employees whose attrition probability falls below a retention threshold in simulation) |
counterfactual: eliminating overtime could potentially retain ≈31 employees (aggregated simulation)
0.09
|
| Main drivers of attrition identified by the model are overtime, business-travel frequency, and promotion opportunities (each having higher influence than salary). Turnover | positive | medium | relative influence of features on predicted attrition probability |
main drivers per model: overtime, business-travel frequency, promotion opportunities (higher influence than salary)
0.09
|
| Local explainability (SHAP and LIME) can identify employee-specific intervention levers for targeted retention actions. Turnover | null_result | high | employee-level change in predicted attrition probability (used to prioritize interventions) |
local explainability (SHAP/LIME) can identify employee-specific intervention levers (for predicted attrition probability)
0.15
|
| The study shifts retention analysis from descriptive correlations and surveys toward actionable, employee-level predictions and policy evaluation. Turnover | null_result | high | operationalization of predictive, actionable attrition estimates (methodological advancement) |
methodological shift: from descriptive correlations to actionable, employee-level predictions and policy evaluation
0.15
|
| Counterfactual simulations show that modest salary increases have a smaller effect on predicted attrition than eliminating overtime (in this dataset and model). Turnover | negative | medium | change in predicted attrition probability and aggregated retained headcount under different simulated interventions |
counterfactuals: modest salary increases have smaller effect on predicted attrition than eliminating overtime (in this dataset/model)
0.09
|
| The IBM HR Analytics dataset was used as a proxy for sustainability-focused (green) roles, relying on objective HR records rather than self-report surveys. Research Productivity | null_result | high | data source / representativeness (proxy use) |
IBM HR Analytics dataset used as proxy for sustainability-focused (green) roles
0.15
|
| Counterfactual simulations reported are predictive rather than causal; estimated effects require causal validation (e.g., randomized trials) before large-scale policy rollout. Research Productivity | null_result | high | validity of counterfactual policy effect estimates (predictive vs causal) |
counterfactual simulations are predictive (not causal); require causal validation before policy rollout
0.15
|
| Generalizability is limited: results based on the IBM dataset may differ for real green-workforce populations, industries, or countries. Research Productivity | null_result | high | external validity / generalizability |
generalizability limited: IBM dataset may not represent real green-workforce populations/industries/countries
0.15
|
| Using calibrated, employee-level predictions enables marginal-cost analyses and prioritization (micro-targeting) to improve retention-efficiency versus uniform, across-the-board policies. Organizational Efficiency | null_result | speculative | potential efficiency gains in retention resource allocation (theoretical outcome) |
calibrated, employee-level predictions enable marginal-cost analyses and prioritization (micro-targeting)
0.01
|
| Economic and organizational benefits (e.g., cost-effective retention, preserved human capital for environmental innovation) are plausible outcomes of applying the approach, but require further causal and cost analyses. Organizational Efficiency | positive | speculative | organizational outcomes: turnover costs avoided, retained human capital, productivity—not empirically measured in the study |
0.01
|
| The paper is the first systematic integration of XAI-based predictive modeling with counterfactual policy simulation specifically targeted at sustainability-oriented HR (Green HRM). Research Productivity | null_result | low | novelty of methodological integration (claim about state-of-the-art) |
claimed novelty: first integration of XAI-based predictive modeling with counterfactual policy simulation in Green HRM literature
0.04
|
| Potential risks of deploying such models include fairness/bias, privacy concerns from employee-level predictions, and adverse morale effects if interventions are unevenly applied. Ai Safety And Ethics | null_result | high | risk categories (fairness, privacy, morale)—qualitative concerns |
identified risks: fairness/bias, privacy concerns, adverse morale from uneven interventions (qualitative)
0.15
|