Support Vector Machines and Kernel Fisher Discriminants: A Case Study using Electronic Nose Data

Kernel methods provide a promising new family of algorithms for ma hine learning and data mining appli ations. In parti ular, kernel-based nonlinear lassi ers su h as support ve tor ma hines (SVMs) and kernel sher dis riminants (KFDs) have been found to work well in pra ti al problems. In addition, there are methods for training these algorithms on large-s ale data sets making them very suitable for use in data mining. In this paper, we evaluate the performan e of SVMs and KFDs on a dataset generated with a ondu ting polymer omposite-based ele troni nose. The ability of SVM and KFD lassi ers to orre tly identify the fun tional lass ( ategory) of a hemi al based on its ele troni nose signature is evaluated and ompared against other more traditional methods, in luding nearest neighbors and linear Fisher dis riminants. Tradeo s between the di erent kernel methods and performan e relative to more traditional methods are dis ussed.