A Selective Approach to Neural Network Ensemble Based on Clustering Technology

A neural network ensemble is a very popular learning paradigm where the outputs of a set of separately trained neural network are combined to form one unified p rediction. To improve the effectiveness of ensemble, neural networks in the ense mble are not only highly correct but make their errors on different parts of the input space as well. However, most existing approaches ensemble all the availab le neural networks for prediction. In this paper, a selective approach to neural network ensemble based on clustering technology is presented. After individual neural networks are trained, the clustering algorithm is used to select a part o f the trained individual networks in order to reduce their similarity. Then many selected neural networks are combined. Experimental results show that this appr oach outperforms the traditional ones that ensemble all of the individual networ ks.