Cross-Silo Federated Learning for Multi-Tier Networks with Vertical and Horizontal Data Partitioning

We consider federated learning in tiered communication networks. Our network model consists of a set silos, each holding vertical partition the data. Each silo contains hub and clients, with silo’s data shard partitioned horizontally across its clients. propose Tiered Decentralized Coordinate Descent (TDCD), communication-efficient decentralized training algorithm for such two-tiered The clients perform multiple local gradient steps before sharing updates their to reduce overhead. adjusts coordinates by averaging workers’ updates, then hubs exchange intermediate one another. present theoretical analysis our show dependence convergence rate on number partitions updates. further validate approach empirically via simulation-based experiments using variety datasets objectives.