A selective approach to neural network ensemble based on clustering technology

Learning for prediction using neural network ensemble can give improved accuracy, and reliable estimation of the generalization error. At present, most approaches ensemble all the available neural networks at hand. In this paper, based on clustering technology, a selective approach to neural network ensemble is presented. After component neural networks are trained, the clustering algorithm is used to select some component neural networks instead of all of the neural networks in order to reduce their similarity. Then selected neural networks are made up to ensemble using simple means method. Finally, an empirical study is conducted and compared with popular ensemble approaches such as bagging. Experimental results show that this approach outperforms the traditional ones that ensemble all of the individual networks.