A time-series classification framework for individual-level absenteeism prediction under severe class imbalance

Workforce management in high-demand sectors like healthcare and emergency services requires proactive staffing to balance operational efficiency with labor costs. Existing machine learning approaches often suffer from a structural limitation: they map features at time t to labels at the same time t, effectively reproducing realized outcomes rather than predicting future events. This paper introduces a Time Series Classification (TSC) framework that separates historical attendance sequences from future absence labels, enabling genuinely proactive prediction.

Because longitudinal attendance data is rarely public, the authors constructed a reproducible simulated dataset calibrated to the UCI 'Absenteeism at Work' benchmark. They evaluated three deep learning architectures—LSTM, CNN, and the hybrid LSTM-FCN—under conditions of severe class imbalance (approximately 97.7% attendance, 2.3% absence). A critical component of the study is the theoretical and empirical analysis of loss functions. The authors compare Binary Focal Loss (BFL) and Geometric Mean (G-Mean) loss, deriving the gradient dynamics to explain why standard BFL parameterizations fail when the majority class is the positive class.

The study establishes that the LSTM-FCN architecture provides the most consistent precision and specificity. Regarding loss functions, the authors prove that standard BFL configurations (designed for minority-positive tasks) amplify majority-class dominance when applied to absenteeism. They derive an analytically balanced weight for BFL, α* = 1/(1+ρ), which significantly improves performance. However, they find that G-Mean loss is superior for practitioners because it exhibits a self-correcting gradient structure that adapts automatically to class imbalance without requiring manual parameter calibration.