Leakage-Aware LLM Augmentation for Attrition Prediction: A DecisionCentric Evaluation

Abstract

Employee attrition is a high-cost, asymmetric decision problem plagued by data imbalance. To address this, we propose a leakage-aware, dual-network augmentation framework that integrates the semantic reasoning of Large Language Models (LLMs) with the distributional rigorousness of GAN discriminators. Specifically, we employ an autoregressive transformer (GReaT) as a generative "actor" to synthesize diverse minority samples, coupled with a post-hoc "critic" derived from a conditional GAN to strictly filter low-fidelity outliers. This actor-critic inspired loop ensures that synthetic records broaden coverage without introducing noise. We evaluate this pipeline using a rigorous protocol where all generation and filtering occur strictly within training folds to prevent leakage. Experiments on HR datasets show that moderate, critic-guided oversampling yields significant recall gains (e.g., raising recall from 45% to 53%) and improved F₁ scores compared to baselines, while maintaining calibration and discrimination (stable ROC-AUC). Furthermore, the approach preserves interpretability (SHAP) and fairness (low TPR gaps), offering HR practitioners a robust, decision-centric tool to identify at-risk employees without sacrificing transparency.