一种神经网络分类器样本数据选择方法

Sample data selection method for neural network classifiers

为了提高神经网络分类器的性能,提出一种基于阴影集的训练样本数据选择方法.在阴影集的基础上提出核数据和边界数据的概念.首先通过模糊C均值聚类(FCM)获取样本数据的最优模糊矩阵;然后诱导出相应的阴影集;样本数据结合阴影集构造核数据和边界数据;最后在核数据和边界数据中进行数据选择.利用该方法,结合Iris数据集分别对BP网络、LVQ网络和可拓神经网络(ENN)等分类器进行实验研究.结果表明:该方法能够保留典型的样本,减少训练样本数据的数量;利用该方法所选择的数据对神经网络分类器进行训练,保证了分类器的泛化能力,节约了训练时间,有效提高分类器的性能.

In order to improve the performance of neural network classifiers （NNCs）, a novel sample data selection method based on shadowed sets was proposed. On the basis of shadowed sets, core data and boundary data were established. First, the optimal fuzzy matrix of sample data was acquired by using FCM. Then, corresponding shadowed sets were induced. On the foundation of sample data and shadowed sets, core data and boundary data could be formed. Finally, the sample data of NNCs could be selected effectively from core data and boundary data. Applying this method and Iris data, experiments for BP neural network, LVQ neural network and extension neural network （ENN） are conducted. Experimental results show that the proposed method can keep typical sample data and reduce the number of training sample data. And with selected sample to train NNCs data can save training time, guarantee generalization ability, and effectively achieve a better performance.