Differential Refresh Policies for Models Trained on Lagging Data Snapshots: From a Single-Age Equivalence Limit to an Optimal Per-Segment Allocation
The paper investigates optimal differential refresh policies for models trained on lagging data snapshots. It shows that while simple uniform timers are equivalent to complex weighted budgets, a differential approach yields better performance gains by allocating refresh budgets based on the square root of segment risk (weight times change rate). This strategy outperforms uniform timing unless the workload is perfectly homogeneous.