@MastersThesis{Son02, author = {S\"oren Sonnenburg}, title = {New Methods for Splice Site Recognition}, school = {Humboldt University}, year = {2002}, note = {supervised by K.-R. M\"uller H.-D. Burkhard and G.~R{\"a}tsch}, ps = {http://sonnenburgs.de/soeren/publications/Son02.ps.gz}, pdf = {http://sonnenburgs.de/soeren/publications/Son02.pdf.gz}, dataset = {http://www.fml.tuebingen.mpg.de/raetsch/projects/AnuSplice/}, abstract = { Modelling \emph{splice sites} is considered a difficult task, and as of this writing, the procedure of splicing is still not well understood. We combine successful \emph{discriminative learners} like Support Vector Machines (SVM) and \emph{descriptive learners} like Hidden Markov Models (HMMs) to separate true splice sites from decoys. Recently developed kernel functions like the TOP- and Fisher Kernel (FK) that are derived from generative models are used to combine SVMs and HMMs. Furthermore, results for the well known Locality Improved Kernel are presented and its connection to the FK, derived from a special HMM is shown. Finally we provide an experimental analysis of splice sites and investigate the classification performance using a variety of learning machines. } }