A0376
Title: fastkqr: A fast algorithm for kernel quantile regression
Authors: Qian Tang - University of Iowa (United States)
Yuwen Gu - University of Connecticut (United States)
Boxiang Wang - University of Iowa (United States) [presenting]
Abstract: Quantile regression is a powerful tool for robust and heterogeneous learning and has been used in a wide spectrum of applied areas. However, its computational cost can be prohibitively high in contemporary large-scale applications due to the nonsmoothness of the quantile loss. A new algorithm named fastkqr is introduced, which provides a significant advance toward computing quantile regression in reproducing kernel Hilbert spaces. The crux of fastkqr is a novel finite smoothing algorithm, which magically gives the exact quantile regression solutions rather than approximations. An interpretability issue of quantile regression, which involves fitted curves crossing at multiple quantile levels in finite samples, is also addressed by presenting a new algorithm for fitting the non-crossing kernel quantile regression by imposing penalizations on the kernel coefficients. Extensive simulations and real applications are used to demonstrate that fastkqr achieves the same accuracy as the state-of-the-art algorithms but can be orders of magnitude faster.
