A novel diversity measure based on geometrical relationship

Abstract

In the complicated pattern recognition, multiple classifier systems (MCSs) can usually obtain higher classification accuracy compared to a single classifier when there is high diversity among member classifiers. Therefore, diversity measures are especially important for the design of MCSs. Most available diversity measures used the consistency or inconsistency of the classification results obtained by member classifiers. Those measures can, to some degree, describe the difference among classifiers, yet not comprehensive and in some cases may cause “diversity submergence”. In this paper a novel geometric relation based diversity measure and a method for MCSs design using the new diversity measure are proposed. It is experimentally shown that the novel diversity measure is rational, which can suppress the “diversity submergence”, and it can be effectively used in designing MCSs.