Labor is typically the largest controllable operating expense in modern enterprises, yet workforce planning remains siloed: demand forecasting, attrition prediction, and headcount budgeting are performed independently, causing systematic over-staffing or costly reactive hiring. This paper presents a unified framework that couples attrition-aware spatio-temporal graph learning with stochastic labor-cost optimization. The organization is modeled as a heterogeneous graph over roles, skills, and departments, on which a spatio-temporal graph transformer jointly forecasts workload demand, role-level staffing requirements, and cohort-level attrition hazards over horizons of 1 to 12 months; a cross-task attention module lets attrition signals recalibrate future capacity. Forecast distributions then feed a two-stage stochastic program that jointly optimizes planned and emergency hiring, overtime, contractor usage, and internal transfers under service-level and budget constraints, yielding auditable plans that are optimal for the sampled scenario set. On a 72-month benchmark for two enterprise-scale organizations (approximately 8,500 and 21,000 employees), synthesized from the public IBM HR Analytics attrition dataset with simulated operational demand, the framework reduces the weighted absolute percentage error (WAPE) of staffing-requirement forecasts by about 15% and 19% relative to independent long short-term memory (LSTM) and Prophet pipelines, and cuts simulated total labor cost by about 6-7% while maintaining a service level of approximately 98%, primarily by shifting spend from emergency hiring and overtime toward planned recruitment; the learned attention also yields faithful, interpretable indicators of team-level attrition risk.
pdf
