1: \begin{abstract}

2: In this paper, we propose a first-order decentralized stochastic optimization method to minimize a finite-sum of smooth and strongly convex cost functions distributed over an arbitrary directed graph. $\AS$ removes the uncertainty caused by the stochastic gradients using node-level variance reduction combined with network-level gradient tracking to manage the data dissimilarity across all nodes. $\AS$ uses row and column stochastic weight matrices to ensure agreement and optimality across the nodes. We show that for a constant step-size,~$\AS$ converges linearly to the exact solution of the global optimization problem. We quantify the directed nature of the underlying graph using a directivity constant~$\psi$, and characterize the  the regimes in which~$\AS$ achieves a linear speed-up over its  centralized counterpart. Numerical experiments illustrate the convergence properties of $\AS$ for both strongly convex and non-convex problems.

3: 

4: \keywords Stochastic optimization, variance reduction, first-order methods, decentralized algorithms, directed graphs

5: \end{abstract}

6: