A theoretical framework for independent classifier combination

Abstract

The combination of classifiers from independent observation domains has a myriad of benefits in practical pattern recognition problems. In this paper we propose a firm theoretical framework from which an upper bound on classifier combination performance can be calculated, based on mismatches between train and test sets. Using this framework, insights can be gained into the conditions under which classifiers can best be combined and where their respective confidence errors stem from. The theoretical framework is presented along with synthetic experiments for empirical validation.