Combined classifier optimisation via feature selection

We propose a new method for performance-constraining the feature selection process as it relates to combined classifiers, and assert that the resulting technique provides an alternative to the more familiar optimisation methodology of weight adjustment. The procedure then broadly involves the prior selection of features via performan-ceconstrained sequential forward selection applied to the classifiers individually, with a subsequent forward selection process applied to the classifiers acting in combination, the selection criterion in the latter case deriving from the combined classification performance. We also provide a number of parallel investigations to indicate the performance enhancement expected of the technique, including an exhaustive weight optimisation procedure of the customary type, as well as an alternative backward selection technique applied to the individually optimised feature sets.