基于核的自组织映射聚类

Kernel-Based Self-Organizing Map Clustering

将核学习的方法应用于自组织映射聚类中,提出了一种核自组织映射聚类算法.该算法以核函数代替原始数据在特征空间中映射值的内积,并且神经元权值向量的初始化和更新都可由其组合系数向量表示,从而获得了直观而简单的迭代公式.分析了算法中学习速率过高会降低学习稳定性、学习速率过低又会降低收敛速度等参数选择问题,给出了一组折中考虑学习稳定性和收敛速度要求的参数初始值.实验结果表明,核自组织映射聚类对于非椭圆型的类分布数据,如环形数据,聚类正确率也能够达到99.886 4%.对IRIS数据集和入侵检测报警数据的聚类也证明了核自组织映射聚类方法的良好性能.

The idea of kernel-based learning method is applied to self-organizing map （SOM） clustering, and an algorithm of kernel self-organizing map （KSOM） clustering is proposed. The inner product of the mapping value of the original data in feature space is replaced by a kernel function, and the weights of each neuron can be initialized and updated by initializing and updating the combinatorial coefficient vector of each weight in the algorithm of KSOM, so some intuitive and simple iteration formulas are obtained. The problems of selecting parameters, such as big learning rate can decrease the learning stability while small learning rate can reduce the convergent speed, were analyzed, and a group of electric values of initial parameters between the learning stability and the convergent speed were yielded. The experimental result shows that the KSOM method can cluster the data with non-spherical shapes such as annular shape, and the cluster precision can reach 99. 886 4%. Examples of clustering IRIS data and alerts in intrusion detection also proved the good performance of the KSOM method.