Joint Research Conference

June 24-26, 2014

Sequential Gaussian Classifiers

Abstract:

Neutral zone classifiers were developed to deal with ambiguity in classification problems. The classifier allows a region of neutrality if there is insufficient evidence to assign an object into one of the groups. Previous work has dealt with single-stage classification decisions. We extend the concept of neutral zone classification to multi-stage contexts, such as longitudinal data paradigms, under the Gaussian distribution assumptions.  Decision boundaries are derived to minimize the expected misclassification cost.  Misclassification rates and expected stopping times are investigated. The results are compared with alternative decision boundaries: repeated single-stage boundaries and truncated sequential probability ratio test boundaries.