Unit measure violations in pattern recognition: Ambiguity and irrelevancy

摘要

Features which are ambiguous or irrelevant for some samples contained in a data array are troublesome to all forms of multivariate analysis. Heretofore, feature redefinition and or excision of subsets of the data array have been the only recourse for the analyst. Those datasets for which these procedures have not been available have often been essentially intractable.In what follows we present a formal structuring of the problem of ambiguity and irrelevancy, both of which introduce non-Bayesian effects into the analytic procedure. A methodology is described for integrating this formalism into a wide variety of pattern recognition techniques.