In this work, we attempt to use an unsupervised learner to construct concepts which we then subsequently use for supervised learning. We do this work within a Minimum Message Length (MML) framework.

The unsupervised learning program which we use is the SNOB program, which is based on MML, although the unsupervised learner need not necessarily be based on MML. We report an empirical study comparing the predictive accuracy of two supervised learners, namely the decision tree programs C4.5 and EFT.

We found that one approach produced a modest improvement in predictive accuracy for two of the six domains discussed, however there was a small decrease in predictive accuracy in some other domains.

Keywords: Machine learning, supervised learning, unsupervised learning, mixture modelling, clustering, instrinsic classification, Minimum Message Length, SNOB.