Induction of category distributions: a framework for classification learning.

We present a framework for classification learning that assumes that learners use presented instances (whether labeled or unlabeled) to infer the density functions of category exemplars over a feature space and that subsequent classification decisions employ a relative likelihood decision rule based on these inferred density functions. A specific model based on this general framework, the category density model, was proposed to account for the induction of normally distributed categories either with or without error correction or provision of labeled instances. The model was implemented as a computer simulation. Results of five experiments indicated that people could learn category distributions not only without error correction, but without knowledge of the number of categories or even that there were categories to be learned. These and other findings dictated a more general learning model that integrated distributional representations based on both parametric descriptions and stored instances.