结合类内集中度和优化RBF神经网络的特征选择

Feature Selection Combined Classificatory Concentration with Improved RBF Neural Network

特征选择是文本分类的一个核心研究课题.首先提出了优化的文档频和类内集中度,紧接着提出了自适应量子粒子群优化算法并用于训练RBF网络的基函数中心和宽度,而且还结合最小二乘法计算网络权值,对RBF神经网络进行了优化,最后提出了一个综合性特征选择方法.该综合性方法首先使用类内集中度过滤掉一些词条以降低文本特征空间的稀疏性,然后再利用优化的RBF网络对特征进行优选.实验结果表明此种特征选择方法有较好的准确率和召回率.

Feature selection is the core research topic in text categorization.Firstly,an optimized document frequency and classificatory concentration were presented.And then,an adaptive quantum-behaved particle swarm optimization（AQPSO） algorithm was provided to improve the performance of radial basis function（RBF） neural network.By applying AQPSO algorithm to train the central position and width of the basis function adopted in the RBF neural network,and computing the weights of the network with least-square method（LSM）,the RBF neural network was improved.Finally,a combined feature selection method was proposed.The combined feature selection method firstly uses classificatory concentration to filter out some terms to reduce the sparsity of feature spaces,and then employs the improved RBF neural network to select more outstanding feature subset.The experimental results show that the combined method is excellent in accuracy rate and recall rate.