A study of all common subsequences in kernel machine

Abstract

Counting all common subsequences (ACS) was proposed as a similarity measurement, which is conceptually different from the sequence kernel (SK) in that ACS only considers the occurrence of subsequences while SK uses the frequency of occurrences of subsequences. This difference evidently results in significant performance variety. ACS has been very competitive in the kNN classifier, however, its performance with kernel machine has been rarely investigated. This is due to the fact that whether ACS is suitable for a kernel classifier is not clear. To this end, this paper firstly proves that ACS is a valid kernel, with a delicate analysis. Then, ACS is further proved to be a good kernel with a comparison with SK in the support vector machine.