RANK AND RESPONSE COMBINATION FROM CONFUSION MATRIX DATA

Abstract

The use of prior behavior of a classifier, as measured by the confusion matrix, can yield useful information for merging multiple classifiers. In particular, response vectors can be estimated and a ranking of possible classes can be produced which can allow Borda type reconciliation methods to be applied. A combination of real data and the simulation of multiple classifiers is used to evaluate this idea, and to compare with eleven other classifier combination techniques. Millions of classifications were used in the evaluation.