A Novel Recursive Kernel-Based Algorithm for Robust Pattern Classification

摘要

Kernel methods comprise a class of machine learning algorithms that utilize Mercer kernels for producing nonlinear versions of conventional linear learning algorithms. This kernelizing approach has been applied, for example, to the famed least mean squares (LMS)algorithm to give rise to the kernel least mean squares (KLMS) algorithm. However, a major drawback of the LMS algorithm (and also of its kernel-ized version) is the performance degradation in scenarios with outliers. Bearing this in mind, we introduce instead a kernel classifier based on the least mean M-estimate (LMM) algorithm which is a robust variant of the LMS algorithm based on M-estimation techniques. The proposed Kernel LMM (KLMM) algorithm is evaluated in pattern classification tasks with outliers using both synthetic and real-world datasets. The obtained results indicate the superiority of the proposed approach over the standard KLMS algorithm.