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.