Federated Optimization of ℓ0-norm Regularized Sparse Learning

Regularized sparse learning with the ℓ0-norm is important in many areas, including statistical and signal processing. Iterative hard thresholding (IHT) methods are state-of-the-art for nonconvex-constrained due to their capability of recovering true support scalability large datasets. The current theoretical analysis IHT assumes use centralized IID data. In realistic large-scale scenarios, however, data distributed, seldom IID, private edge computing devices at local level. Consequently, it required study property a federated environment, where update model individually communicate central server aggregation infrequently without sharing this paper, we propose first group methods: Federated Hard Thresholding (Fed-HT) (FedIter-HT) guarantees. We prove that both algorithms have linear convergence rate guarantee optimal estimator, which comparable classic methods, but decentralized, non-IID, unbalanced Empirical results demonstrate Fed-HT FedIter-HT outperform competitor—a distributed IHT, terms reducing objective values fewer communication rounds bandwidth requirements.