Exploiting Generative Models in Discriminative Classifiers-英文文献.pdf

Exploiting generative mo dels in discriminative classiers Tommi S Jaakkola and David Haussler ftommihaussler gcseucs cedu Department of Computer Science University of California Santa Cruz CA and Isaac Newton Institute for Mathematical Sciences University of Cambridge Clarkson Road Cambridge CB EH UK Abstract Generative probability mo dels such as hidden Markov mo dels pro vide a principled way of treating missing information and dealing with variable length sequences On the other hand discriminative metho ds such as supp ort vector machines enable us to construct exible decision b oundaries and often result in classication p er formance sup erior to that of the mo del based approaches An ideal classier should combine these two complementary approaches In this pap er we develop a natural way of achieving this combina tion by deriving kernel functions for use in discriminative metho ds such as supp ort vector machines from generative probability mo d els We provide a theoretical justication for this combination as well as demonstrate a substantial improvement in the classication p erformance in the context of DNA and protein sequence analysis Intro duction Sp eech vision text and biosequence data can b e dicult to deal with in the context of simple statistical classication problems Because the examples to b e classied are often sequences or arrays of variable size that may have b een distorted in par ticular w