A neural network ensemble method with new definition of diversity based on output error curve
Yang Yang, Yuan Feng Xu, Qun‐Xiong Zhu · 2010
Neural network ensemble can significantly improve generalization accuracy of networks by training several networks and combining their results. The traditional way to define diversity only considers the inner structure of networks. However, because neural network is a “black box”, it is blind to search for diversity through network structure. This paper proposed a method to find diversity from the output error curves of neural networks, it only pays attention to output error curve and thus avoids involving inner structure of neural networks. We apply this method on several dataset including UCI dataset and a practical industrial dataset. The results indicate the effectiveness of this method. This method can be extended to not only neural network ensemble but also multiple-model ensemble which provides new thoughts for the development of neural network ensemble.