Abstract Quantile regression has become a popular alternative to least squares regression for providing a comprehensive description of the response distribution, and robustness against heavy-tailed error distributions. However, the nonsmooth quantile loss poses new challenges to distributed estimation in both computation and theoretical development. To address this challenge, we use a convolution-type smoothing approach and its Taylor expression to transform the nondifferentiable quantile loss function into a convex quadratic loss function, which admits a fast and scalable algorithm to perform optimization under massive and high-dimensional data. The proposed distributed estimators are both computationally and communication efficient. Moreover, only the gradient information is communicated at each iteration. Theoretically, we show that, after a certain number of iterations, the resulting estimator is statistically as efficient as the global estimator without any restriction on the number of machines. Both simulations and data analysis are conducted to illustrate the finite sample performance of the proposed methods.
Paper
Full text
Smoothing quantile regression for a distributed system
Semantic Scholar · Mathematics · 2021
Abstract
Abstract Quantile regression has become a popular alternative to least squares regression for providing a comprehensive description of the response distribution, and robustness against heavy-tailed error distributions. However, the nonsmooth quantile loss poses new challenges to distributed estimation in both computation and theoretical development. To address this challenge, we use a convolution-type smoothing approach and its Taylor expression to transform the nondifferentiable quantile loss function into a convex quadratic loss function, which admits a fast and scalable algorithm to perform optimization under massive and high-dimensional data. The proposed distributed estimators are both computationally and communication efficient. Moreover, only the gradient information is communicated at each iteration. Theoretically, we show that, after a certain number of iterations, the resulting estimator is statistically as efficient as the global estimator without any restriction on the number of machines. Both simulations and data analysis are conducted to illustrate the finite sample performance of the proposed methods.