Abstract
In recent years, many distributed algorithms towards big data quantile regression have been proposed. However, they all rely on the data are stored in random manner. This is seldom in practice, and the violation of this assumption can seriously degrade their performance. Moreover, the non-smooth quantile loss brings inconvenience in both computation and theory. To solve these issues, we first propose a convex and smooth quantile loss, which converges to the quantile loss uniformly. Then a novel pilot sample surrogate smooth quantile loss is constructed, which can realize communication-efficient distributed quantile regression, and overcomes the non-randomly distributed nature of big data. In theory, the estimation consistency and asymptotic normality of the resulting distributed estimator are established. The theoretical results guarantee that the new method is adaptive to the situation where the data are stored in any arbitrary way, and can work well just as all the data were pooled on a single machine. Numerical experiments on both synthetic and real data verify the good performance of the new method.
•Convex and smooth quantile regression loss is constructed.•Distributed inference for non-randomly stored big data is proposed.•The new method can be easily implemented, and is communication-efficient.•Theoretical properties are established under mild conditions.