Distributed sparse composite quantile regression in ultrahigh dimensions

We examine distributed estimation and support recovery for ultrahigh dimensional linear regression models under a potentially arbitrary noise distribution. The composite quantile regression is an efficient alternative to the least squares method, and provides robustness against heavy-tailed noise while maintaining reasonable efficiency in the case of light-tailed noise. The highly nonsmooth nature of the composite quantile regression loss poses challenges to both the theoretical and the computational development in an ultrahigh-dimensional distributed estimation setting. Thus, we cast the composite quantile regression into the least squares framework, and propose a distributed algorithm based on an approximate Newton method. This algorithm is efficient in terms of both computation and communication, and requires only gradient information to be communicated between the machines. We show that the resultant distributed estimator attains a near-oracle rate after a constant number of communications, and provide theoretical guarantees for its estimation and support recovery accuracy. Extensive experiments demonstrate the competitive empirical performance of our algorithm.