Graph-Regularized Survival Learning with Optimal Transport for Individual Employee Turnover Forecasting

This study forecasts attrition or job-change risk for individual employee records; department- and city-level quantities are secondary aggregates of those predictions. We propose Graph-Regularized Survival Learning with Optimal Transport (GRSOT), which combines a survival-style risk objective, entropic optimal transport, and Laplacian regularization over a static employee-similarity graph. The general formulation accepts event times and censoring indicators, whereas the two public benchmarks used here are cross-sectional and provide neither longitudinal employment trajectories nor observed interaction networks. We therefore construct proxy time/event pairs from tenure-like variables and binary labels, and interpret the empirical results as survival-inspired risk forecasting rather than evidence from genuine longitudinal histories. The graph performs similarity-based risk smoothing and does not represent causal peer influence. The transport module targets covariate shift between controlled source and target partitions and assumes that the conditional proxy outcome mechanism remains stable. Relative to OT-Surv, GRSOT increases the C-index from 0.812 to 0.836 and reduces expected calibration error from 0.045 to 0.037 on IBM-HR; on HR-JC, the corresponding values improve from 0.781 to 0.803 and from 0.053 to 0.045. Training adds a sparse graph linear solve and Sinkhorn scaling, while deployment requires one encoder pass, local graph smoothing, and baseline-hazard evaluation. The resulting outputs support individual risk ranking, proxy-horizon calibration, and non-causal group-level decomposition for decision support.