Output details
13 - Electrical and Electronic Engineering, Metallurgy and Materials
University of Surrey
Minimising Added Classification Error Using Walsh Coefficients
Two-class supervised learning for classifier ensembles is formulated as learning an incompletely specified Boolean function. This research uses Walsh spectral coefficients computed from the Boolean function to select optimal complexity of the base classifiers, using only the training set. State-of-the-art is furthered by postulating a theoretical understanding of this design process using a well-known model. The significance is that time-consuming cross-validation is not required to set classifier parameters. Supported by grant EPSRC E061664/1 £260,000, and the approach is incorporated into EU project PIEF-GA-2009-254451, the aim of which is to optimise the accuracy/diversity trade-off in Multiple Classifier Systems.