Dynamic Alignment Kernels

Chris Watkins · The MIT Press eBooks · 2000

There is much current interest in kernel methods for classification, regression, PCA, and other linear methods of data analysis. Kernel methods may be particularly valuable for problems in which the input data is not readily described by explicit feature vectors. One such problem is where input data consists of symbol-sequences of different lengths, and the relationships between sequences are best captured by dynamic alignment scores. This paper shows that the scores produced by certain dynamic alignment algorithms for sequences are in fact valid kernel functions. This is proved by expressing the alignment scores explicitly as dot-products. Alignment kernels are potentially applicable to biological sequence data, speech data, and time series data. The kernel construction may be extended from pair HMMs to pair probabilistic context-free grammars. Introduction: Linear Methods using Kernel Functions 1 1 Introduction: Linear Methods using Kernel Functions In many types of machine learni...

Cited by 244 works. doi:10.7551/mitpress/1113.003.0006

Authors