基于改进ART2算法的数据聚类方法

Data Clustering Based on Improved Algorithm of ART2

通过对经典自适应谐振(Adaptive Resonance Theory,ART)神经网络聚类过程的分析,指出其存在警戒参数主观设置,过分依赖获胜神经元信息,输出无层次结构等不足之处.在此基础上,提出基于ART2神经网络的改进聚类算法.本算法通过在竞争过程中同时考虑获胜神经元和其它神经元的信息以及Hebb规则来实现通过单个ART神经网络的多层动态聚类结构(对于一定范围粒度的聚类也不再需要重新训练神经网络),除此还降低了对警戒参数主观设置的要求.这些优点有效满足聚类的基本要求,能够避免采用级联结构实现层次聚类带来的性能、参数设置等问题.

By analyzing the clustering process of classical Adaptive Resonance Theory Network （ART2）, this paper points out the shortcomings of ART2： subjective setting of vigilance parameter, excessive dependence on winning neuron information and output without hierarchical structure, etc. So an improved clustering algorithm of ART2 has been presented. This algorithm can realize multi-layer dynamic clustering structure through a single ART2 neural network by simultaneously taking into consideration the information from both the winning neuron and other neurons in process of competition as well as Hebb rule. Therefore,there is no demand of retraining neural network within a certain range of granularity. Besides, this algorithm reduces the requirement of subjective setting of vigilance parameter. The above mentioned advantages can effectively satisfy fundamental demands of clustering and get rid of the problems of performanee and parameter setting caused by employing cascade structure to realize hierarchical clustering structure.