A novel measure method for diversity of classifier integrations using complement information entropy

A novel diversity measure method using complement information entropy (CIE) is proposed to solve the problem that the diversity estimation of multiple classifier systems is unable to deal directly with fuzzy data. A set of base classifiers is generated by using training data, and then is used to label test data. The outputs of the classifiers are reorganized into a new classification data space. Then the complement information entropy model is introduced under fuzzy relation to measure uncertainty information of the new space and the uncertainty information is used to estimate the diversity of base classifiers. Finally, an ensemble system is constructed based on the criterion that the ensemble diversity of the classifier set increases when a base classifier is added, and the ensemble system is used to validate the performance of CIE. Experimental results and a comparison with the Q-statistic method show that the average classification accuracy of CIE increases by 2.03%, and the number of ensemble classifiers reduces by 17%. Moreover, CIE also improves the ability of ensemble systems to process diverse data.