A classifier combination tree algorithm

In recent years a number of authors have suggested that combining classifiers within local regions of the measurement space might yield superior classification performance to rigid global weighting schemes. In this paper we describe a modified version of the CART algorithm, called ARPACC, that performs local classifier combination. One obstacle to such combination is the fact that the 'optimal' covariance combination results originally assumed only two classes and classifier unbiasedness. In this paper we adopt an approach based on minimizing the Brier score and introduce a generalized matrix inverse solution for use in cases where the error matrix is singular. We also report some preliminary experimental results on simulated data.