一种快速、鲁棒的有限高斯混合模型聚类算法

摘要

Finite mixture model is an effective clustering method based on probability model.Aiming at the clustering algorithm of Gaussian mixture model This paper imposed entropy penalized operators on the mixed coefficients of components and the labels of samples respectively,which brings to two levels controls for the number of components and rapid reduction of the illegitimate ones.Thus the algorithm converges to exact solutions with only a few iterations.Since the traditional algorithm is very sensitive to the initial values (for example,the number of components must be set in advance),which often leads to the EM algorithm to fall into local optima or converges to the boundary of the solution space,the new algorithm of this paper is very robust and has no special demands for the initializations,just testified by the experiments.%有限混合模型聚类是一种基于概率模型的有效聚类方法.针对高斯混合模型的聚类算法,分别对模型的成分混合系数及样本所属成分的概率系数施加熵惩罚算子,实现对模型成分数的两级控制,快速消除无效成分,使算法能在很少的迭代次数内收敛到确定解.传统算法对初始值(成分数目c需事先指定)的设置非常敏感,容易导致EM算法陷入局部最优解或收敛到解空间的边界,而文中的算法对初始值的设定没有特殊的要求,实验证明其具有很好的鲁棒性.