一种两阶段的神经网络属性选择方法

A TWO-PHASE METHOD OF NEURAL NETWORK FEATURE SELECTION

神经网络的输入属性选择一直是一个比较困难的问题.由于神经网络反复训练的时间复杂度,Wrap-per方法是不适用的,而单纯使用Filter方法也难以获得很好的分类精度.文中提出了一种两阶段的神经网络属性选择方法,以综合Filter和Wrapper两类方法的优势.该方法首先采用基于不一致率的遗传算法GFSIC来删除属性集合中的无关属性,然后采用基于敏感性度量的属性选择算法SBFCV来删除冗余和无用的属性.研究和实验结果表明,该方法可以有效地删除原始数据中的无关和冗余属性,增强神经网络的泛化能力.

Neural network feature selection is an open issue hard to solve.Because of the time complexity of retraining the network,wrapper methods are infeasible.On the other hand,filter methods are not enough to get good classification accuracy.This paper presents a twophase feature selection method,which takes advantage of both filter and wrapper methods.It begins the first phase by running GFSIC algorithm,a filter method based on inconsistency,to remove irrelevant features.In the following phase,it runs SBFCV algorithm,a wrapper method based on sensitivity,to remove redundant or useless features from neural network's consideration.Analysis and experimental studies show that the new method can perform feature selection effectively and improve the generalization of neural network as well.