MLP Based Linear Fea ure Ex rac ion for Nonlinearly Separable Data

Abstract:

A novel approach to linear feature extraction is presented. Most supervised feature extraction algorithms use mean square error or other measures based on the difference between expected and actual output values as a performance criterion. The novel approach presented here uses data visualisation together with an empirical classification error (percentage of cases classified incorrectly) as performance criterion. To find the optimal data transformation weights, the Multilayer Perceptron cost function with a special regularisation term is applied. The technique proposed is verified and compared with five competing mapping techniques with respect to visualisation and different classification error criteria. For comparison, two artificial and 12 real world data sets are used.