MxBoost: Mutual-exclusive Boosting for Online Classification

摘要

In this paper we discuss how to improve classifier ensembles for online classification problems where classification has to be performed quickly. Under the restriction on time and therefore ensemble size, we propose MxBoost (Mutual-exclusive Boosting) based on the heuristic that members of a decision committee should make mistakes mutually exclusively. Theoretical justifications are given to explain the effectiveness of MxBoost. Furthermore, we conduct extensive empirical studies using representative ensemble schemes including One Per Class (OPC) and Error Correcting Output Coding (ECOC). Our experiments show that MxBoost improves the class-prediction accuracy of both OPC and ECOC on all standard and image datasets that we tested while maintaining good runtime performance. Our experiments also show that MxBoost does not suffer from the overfilling problem.