Stochastic Douglas-Rachford Splitting for Regularized Empirical Risk Minimization: Convergence, Mini-batch, and Implementation

21 Sept 2022, 19:24 (modified: 22 Nov 2022, 16:44)Accepted by TMLREveryoneRevisionsBibTeX
Abstract: In this paper, we study the stochastic Douglas-Rachford splitting (SDRS) for general empirical risk minimization (ERM) problems with regularization. Our first contribution is to prove its convergence for both convex and strongly convex problems; the convergence rates are $O(1/\sqrt{t})$ and $O(1/t)$, respectively. Since SDRS reduces to the stochastic proximal point algorithm (SPPA) when there is no regularization, it is pleasing to see the result matches that of SPPA, under the same mild conditions. We also propose the mini-batch version of SDRS that handles multiple samples simultaneously while maintaining the same efficiency as that of a single one, which is not a straight-forward extension in the context of stochastic proximal algorithms. We show that the mini-batch SDRS again enjoys the same convergence rate. Furthermore, we demonstrate that, for some of the canonical regularized ERM problems, each iteration of SDRS can be efficiently calculated either in closed form or in close to closed form via bisection---the resulting complexity is identical to, for example, the stochastic (sub)gradient method. Experiments on real data demonstrate its effectiveness in terms of convergence compared to SGD and its variants.
License: Creative Commons Attribution 4.0 International (CC BY 4.0)
Submission Length: Regular submission (no more than 12 pages of main content)
Changes Since Last Submission: None. We just changed the format to the accepted form.
Assigned Action Editor: ~Robert_M._Gower1
Submission Number: 452