Margins and combined classifiers
Llew Mason · 1999
Many binary classification algorithms produce real-valued predictions which are then thresholded to produce a binary classification. These real-valued predictions can often be viewed as a measure of confidence in classification. Recent techniques for the theoretical analysis of such algorithms in terms of the classification confidence (or margin) have been dubbed margins analysis. This form of analysis has been shown to often yield more refined measures of classifier complexity and more realistic measures of algorithm performance than is provided by the now classical VC theory. This thesis is concerned with the application of margins analysis to classifiers which can be represented as thresholded convex combinations.