Supervised clustering using decision trees and decision graphs: An ecological comparison.

Abstract

n this paper, we outline some of the problems in computer learning, particularly with respect to decision trees. We then consider how, in some cases, a decision graph may provide a solution to some of these problems. We compare a decision graph analysis with a decision tree analysis of salt marsh data, predicting predetermined vegetation types from environmental properties. All analyses use a minimum message length criterion to select an optimal model within a class, thereby avoiding subjective decisions. Minimum message length also provides a criterion for choosing between the model classes of tree and graph. In addition to the computational evaluation of models, we examine the ecological implications of the selected solutions. Even if sub-optimal, it is possible that a result can contribute to understanding of the underlying real system.