Communication Efficient Coresets for Empirical Loss Minimization

In this paper, we study the problem of empirical loss minimization with l2-regularization in distributed settings with significant communication cost. Stochastic gradient descent (SGD) and its variants are popular techniques for solving these problems in large-scale applications. However, the communication cost of these techniques is usually high, thus leading to considerable performance degradation. We introduce a novel approach to reduce the communication cost while retaining good convergence properties. The key to our approach is the construction of a small summary of the data, called coreset, at each iteration and solve an easy optimization problem based on the coreset. We present a general framework for analyzing coreset-based optimization and provide interesting insights into existing algorithms from this perspective. We then propose a new coreset construction and provide its convergence analysis for a wide class of problems that include logistic regression and support vector machines. Preliminary experiments show encouraging results for our algorithm on real-world datasets.