Quaternion kernel partial least squares regression algorithms

Abstract

This work provides three quaternion kernel partial least squares (PLS) algorithms for linear and nonlinear regressions. Firstly, the problem of large ill-conditioned matrices is tackled and two specifically designed linear kernel algorithms are suggested. Secondly, since PLS can present low regression accuracy and prediction performance for nonlinear data, a kernel algorithm for performing quaternion nonlinear regression is also given. Computational results and discussion illustrate the relative merits of the algorithms proposed over closely related regression methods.