Linear classification

The chapter focuses on issues related to adopting parametric regression methods to the classification task. This is essentially based on using a composite model representation function, consisting of a real‐valued inner representation function and a discrete outer representation function that assigns class labels based on the former. In principle, parametric model representation is applicable to both classification and regression models, since the employed representation function can be real valued or discrete valued. Parameter estimation can be viewed as an optimization process in which the space of possible parameter vectors is searched for the one that optimizes an adopted performance measure. It is common for classification tasks to involve discrete attributes, either alone or along with continuous attributes. For parametric classification to be applicable to such tasks the inner representation function and its gradient must be capable of handling discrete attributes.