Greedy optimization classifiers ensemble based on diversity

摘要: Decreasing the individual error and increasing diversity among classifiers are two crucial factors for improving ensemble performances. Nevertheless, ''kappa-error'' diagram shows that enhancing is at expense of reducing accuracy. Hence, a new method named Matching Pursuit Optimization Ensemble Classifiers (MPOEC) proposed in this paper order to balance MPOEC adopts greedy iterative algorithm matching pursuit search an optimal combination entire classifiers, eliminates some similar or poor by giving zero coefficients. In approach, coefficient every classifier gained minimizing residual between target function linear basis functions, especially, when functions similar, their coefficients will be close zeros one iteration optimization process, which indicates obtained based on individuals. Because given coefficients, approach may also considered as selective method. Experimental results show improves performance compared with other methods. Furthermore, kappa-error diagrams indicate increased standard strategies evolutionary ensemble.