A hierarchical classifier based on adaptive resonance theory

Abstract

An artificial neural network model is developed that performs the task of hierarchical classification. The model, HAM, is composed of modules based on a simplified version of ARTMAP and is developed over three iterations. The performance and behavior of HAM is investigated at each stage and a large amount of experimental data is provided. Prediction accuracy is used as the primary metric for performance, with the number of learned categories and the general shape of the learned classification hierarchies also being considered. The first iteration, HAM(1), is a supervised version of the ART based hierarchical clustering network HART-S. It allows an off-line training mode and produces very large classification hierarchies with prediction accuracy worse than ARTMAP. The second iteration of the model, HAM(2), improves prediction accuracy and generalization and also allows learning to take place using the on-line training mode. The final version of HAM produces improved classification hierarchies in terms of compactness, shape, and prediction accuracy on the two benchmark datasets. For both on-line and off-line training modes the prediction accuracy of HAM is better than ARTMAP. An additional contribution of this thesis is an input encoding method that improves the prediction accuracy of both ARTMAP and HAM while simultaneously reducing the number of categories needing to be learned.