A New Ensemble Learning Algorithm Based on Improved K-Means for Training Neural Network Ensembles

The diversity of individual neural network will affect the forecast error of neural network ensemble. In this paper, a new training method for neural network ensemble which is based on the improved K-means algorithm is presented. The space diversity among the sample datasets can be realized by clustering the entire dataset using K-means algorithm. To avoiding the sample subset too simple, the sample subsets are interpolated some samples which are randomly chosen from the entire sample space. By contrast experiments with Boosting and Bagging, the presented learning algorithm is proved that can reduce the prediction error of neural network ensemble and enhance the prediction accuracy.