Communication-Efficient Variance-Reduced Decentralized Stochastic Optimization Over Time-Varying Directed Graphs

In this article, we consider the problem of decentralized optimization over time-varying directed networks. The network nodes can access only their local objectives, and aim to collaboratively minimize a global function by exchanging messages with neighbors. Leveraging sparsification, gradient tracking, variance reduction, propose novel communication-efficient scheme that is suitable for resource-constrained We prove in case smooth strongly convex objective functions, proposed achieves an accelerated linear convergence rate. To our knowledge, first framework networks such rate applies settings requiring sparsified communication. Experimental results on both synthetic real datasets verify theoretical demonstrate efficacy scheme.