Modular Neural Network based on Biased Selection of Boundary Region Samples

In view that the traditional modular neural network did not consider the bias of the samples in the cross-boundary region to the cluster attribution when dividing the clustering module, which leads to the data sample hybridity and the poor performance of the final neural network model. This paper presents a modular neural network based on the biased selection of samples in the boundary region. By combining the two-stage clustering of Canopy and fuzzy rough C-means clustering, the number of clusters is adaptively determined, and then the modules are divided. At the same time, taking advantage of the rough-set in uncertain information processing, the samples in the lower approximation set and cross-boundary region are studied separately, and the bias threshold is introduced to further divide the samples in the boundary region. The proposed modular neural network can make the new input sample find the sub network more accurately. The experimental results show that the modular neural network based on the biased selection of samples in the boundary region has obvious advantages in the prediction accuracy and the generalization performance.