模糊粗糙神经网络特征选择方法研究

Fuzzy-rough Neural Network Feature Selection

实际采集的数据中往往存在模糊不确定性和粗糙不确定性,为全面度量数据的不确定性,引入了模糊粗糙集中的模糊粗糙隶属函数概念,并结合容错能力较强的神经网络设计了一种新的模糊粗糙神经网络.该网络不仅训练速度快,且具有较好的分类效果.利用该网络设计了一种新的特征选择算法,根据精度下降指标对输入节点进行结构修剪,实现最优特征子集的搜索.通过UC I数据集实验,并与RBF网络选择结果进行比较,表明该算法具有精度高、速度快、泛化性能好等优点,是有效的.

For the sake of measuring fuzzy uncertainty and rough uncertainty of real datasets, the fuzzy-rough membership function （FRMF） defined in fuzzy-rough set is introduced. A new fuzzy-rough neural network （FRNN） is constructed based on neural network implementation of FRMF. FRNN has the merits of quick learning and good classification performance. A new neural network feature selection algorithm based on FRNN is designed. The input nodes are pruned according to the descent of accuracies and at the same time the search of optimal feature subset is realized. The test results on UCI datasets show that the algorithm is quick and effective, and has better selection precision and generalization ability than RBF feature selection.