神经模糊网络特征选择

Feature Selection Based on Neuro-Fuzzy Networks

基于人工神经网络的特征选择算法一般可以看作是剪枝算法的一个特例:通过剪枝输入节点,计算网络输出对该输入节点对应特征的敏感性。但这些方法往往要求首先对数据做归一化的工作,这可能会改变原数据具备的对分类很重要的某些性质,神经模糊网络是具有自学习能力的模糊推理系统,本文将其与基于隶属度空间的剪枝技术结合起来提出新的特征选择算法,其特点是隶属度函数是自适应学习的,且学习过程在特征选择之前完成,分别对自然数据和人工数据进行实验,并与其它方法相比,结果证明该算法是有效的。

To some extent, the feature selection algorithms based on artificial neural networks can be regarded as the special cases of the architecture pruning algorithms . However , they usually require preprocessing of data normalization, which may change the distribution of the original data which is important to the classification. Neuro-fuzzy networks are fuzzy inference systems with self-study ability. In this paper it is combined with the architecture pruning algorithm based on membership space and a new feature selection algorithm is proposed. The membership functions of the algorithm are learned adaptively , and the learning process is finished before the feature selection. Experiments on natural and synthesized data are given and compared with some traditional techniques. The results show that the proposed method is superior to the traditional ones.