A composite classifier system design: Concepts and methodology

This study explores the scope for achieving enhanced recognition system performance through deployment of a composite classifier system consisting of two or more component classifiers which belong to different categories. The domains of deployment of these individual components (classifiers) are determined by optimal partitioning of the problem space. The criterion for such optimal partitioning is determined in each case by the characteristics of the classifier components. An example, in terms of partitioning the feature space for optimal deployment of a composite system consisting of the linear and nearest neighbor (NN) classifiers as its components, is presented to illustrate the concepts, the associated methodology, and the possible benefits one could expect through such composite classifier system design. Here, the optimality of the partitioning is dictated by the linear class separability limitation of the linear classifier and the computational demand characteristics of the NN classifier. Accordingly, the criterion for the optimal feature space partitioning is set to be the minimization of the domain of application of the NN classifier, subject to the constraint that the linear classifier is to be deployed only in regions satisfying the underlying assumption of linear separability of classes. While many alternatives are available for the solution of the resulting constrained optimization problem, a specific technique-Sequential Weight Increasing Factor Technique (SWIFT)- was employed here for convenience in view of previous successful experience with this technique in other application areas. Numerical results derived using the well-known IRIS data set are furnished to demonstrate the effectiveness of the new concepts and methodology.