An Effective Martin Kernel for Time Series Classification

摘要

Time series classification has attracted a lot of attention in recent years. However, the original data often corrupted with noise. To alleviate this problem, many approaches try to perform nonlinear transformation, such that the resulting space could give out the most relevant features. Since the resulting space is not a Euclidean space, strong assumptions are needed for many kernel-based methods for the purpose of obtaining a reasonable measurement. In this paper we propose a novel approach based on Martin distance. The Martin distance is applied to measure the pairwise distance in the resulting space, without imposing strong assumptions on model states. Experiments on several benchmark datasets demonstrate the advantages of the proposed kernel on its effectiveness and performance.