NWFNN : A Clustering Algorithm Based on Fuzzy Neural Network with Its Application in Text Mining

摘要

The limitation of FCM algorithm is applicable to spheroid and the main defect of traditional methods of fuzzy neural network and fuzzy clustering algorithm is to know the number of clustering in advance. A new clustering algorithm NWFNN is presented in the paper. Euclidean distance based on feature weight is used to non-spherical data space. NWFNN algorithm applies average information entropy to find the number of clusters and a dopts a density function algorithm to find the initial cluster centers. NWFNN uses the fuzzy central vectors as the weights of the neuron network. The winner unit in the model is acquired by comparing the membership degree values between neurons. Only two neurons in the network are updated. The weights of the neuron with the largest membership degree values are updated by a larger learning rate, while the weights of the neuron with the second largest membership degree value are updated by a smaller learning rate, and the weights of other neurons are kept invariable. According to the formula of FCC algorithm, both the fuzzy center clustering vectors (weights of neuron network) and the membership degrees are adjusted, and the number of clustering is determined after the neuron network reaches stable, NWFNN algorithm is applied to text mining. Compared with the traditional fuzzy neuron networks, the present model possesses the simpler structure, higher precision and the higher efficiency, and overcomes the defect that the traditional algorithms need to know the number of clustering in advance. An example demonstrates the effectiveness of the present algorithm.