Flexible decision trees in a general data-mining environment

Abstract

We describe a new data-mining platform, CDMS, aimed at the streamlined development, comparison and application of machine learning tools. We discuss its type system, focussing on the treatment of statistical models as first-class values. This allows rapid construction of composite models - complex models built from simpler ones - such as mixture models, Bayesian networks and decision trees. We illustrate this with a flexible decision tree tool for CDMS which rather than being limited to discrete target attributes, can model any kind of data using arbitrary probability distributions. © Springer-Verlag 2003.