1: \begin{abstract}

2: In this paper, we will focus on solving the splitting problem which is minimizing the sum of two convex functions subject to a linear constraint. This problem has attracted tremendous attention because of its wide applications to machine learning problems, such as Lasso, group Lasso, Logistic regression and image processing. A recent paper by Gu et al (2015) developed a Semi-Proximal-Based Strictly Contractive Peaceman-Rachford Splitting Method (SPB-SCPRSM), which is an extension of Strictly Contractive Peaceman-Rachford Splitting Method (SCPRSM) proposed by He et al (2014). By introducing semi-proximal terms and using two different relaxation factors, SPB-SCPRSM showed a more flexiable applicability compared with its origin SCPRSM and widely-used Alternating Direction Method of Multipliers (ADMM) algorithm, though all of them have $O(1/t)$ convergence rate. 

3: %However, SPB-SPRSM cannot scale to problems with huge number of samples since each single update uses the whole training data set.%

4: In this paper, we develop a stochastic version of SPB-SCPRSM, where only a subset of samples (or even only one sample) are used at each iteration.

5: The resulting algorithm (Stochastic SPB-SCPRSM) can not only scale to problems with huge number of samples and also be more flexiable than Stochastic ADMM on the numerical experiment by setting semi-proximal terms and relaxation factors. Moreover, we show that our proposed method has $O(1/\sqrt{t})$ convergence rate in ergodic sense in general and $O(\log(t)/t)$ in strong convexity case.

6: % But as shown in He et al (2014) that SPRSM based algorithms will always converge faster than ADMM in apllication, our proposed algorithm will also preserve this advantage.

7: %  

8: \end{abstract}

9: