An Explainable Generative AI Framework for Multi-Criteria Employee Churn Analysis

Full abstract

Read the full abstract

Abstract Multi-Criteria Decision-Making (MCDM) approaches often face practical challenges, including inconsistent assessments, variability in judgments, and difficulties in integrating multiple evaluation criteria. This study proposes an AI-driven decision-support framework to address these challenges in employee churn evaluation. A Conditional Wasserstein GAN with Gradient Penalty (CWGAN-GP) is trained on employee data to model relationships between employee features and multiple evaluation indicators. Latent noise vectors are incorporated during generation to introduce controlled stochastic variability, allowing the model to produce multiple evaluation samples for the same employee profile while preserving statistical consistency with the learned data distribution. The generated evaluation scores across several decision criteria are aggregated using TOPSIS to rank employees and classify them into three churn-risk categories. To enhance computational efficiency, machine learning classifiers are subsequently trained as surrogate models to approximate the TOPSIS-derived churn-risk categorization for new employee profiles. Explainability is provided through SHAP-based global and local interpretations, supported by counterfactual explanations. Overall, the proposed framework integrates generative modeling, multi-criteria decision analysis, and explainable AI as a methodological decision-support framework for structured employee churn-risk analysis.