Adaptive Data Admission and Retention for Streaming Federated Learning

This paper studies streaming federated learning when clients have limited memory, newly generated samples have time-varying acquisition costs, and retained buffers must satisfy capacity constraints. It derives a learning-error bound based on effective sample size, accounting for instantaneous sample counts, distinct-sample growth, and reuse imbalance. The proposed Active-Constraint Drift-Plus-Penalty (ACDPP) policy combines a structured client-side K-step retention rule with server-side online admission and a time-varying rectangular admission region. The authors provide sublinear regret and sampling-cost violation guarantees, while controlling buffer occupancy through offline selection of the retention horizon. Experiments on multiple datasets report performance close to an oracle benchmark under the stated constraints.