Some analytical results on critic-driven ensemble classification

We (1999) proposed a framework for ensemble classification wherein auxiliary networks, dubbed critics, are used to provide reliability information on the ensemble's individual classifiers/experts. We showed experimentally that critic-driven combining schemes extend the applicability of ensemble methods by overcoming the usual requirement that the individual classifier error rate p must be less than 0.5. Here, we support our previous work by proving, under an independence assumption, that performance for a particular critic-driven voting scheme improves with increasing ensemble size N, so long as p+q<1, with p the critic's error rate in predicting accuracy of expert decisions. While this independence analysis gives significant insight into the conditions for success of critic-based schemes, it does not accurately predict the ensemble performance curve. We thus also develop an analytical approach for predicting the curve, by modeling dependence between experts.