Bayesian classification methods

Suppose you are handed a collection of cases with no labels attached, and the job is to decide which class each one belongs to. The nature of that job depends on what you already know. When the full roster of possible classes is laid out ahead of time, you are doing supervised classification. When the classes themselves are unknown and have to be discovered from the data, the problem turns into unsupervised classification instead.

This paper surveys a range of Bayesian ways of attacking both versions. On the supervised side sit methods such as discrimination; on the unsupervised side, techniques like clustering. For the supervised case the discussion goes a step further and works out a fully Bayesian theoretical rule, one that can act as the foundation from which particular model-based classification rules are constructed.

The topics touched along the way include Bayesian decision theory, cluster analysis, configural frequency analysis, discriminant analysis, logistic regression, and statistical learning.