一种具有抗噪能力的贝叶斯可能性聚类方法  

Bayesian possibility clustering method with anti-noise ability

作　　者：张聪[1] 顾晓清[1] 王洪元[1] Zhang Cong;Gu Xiaoqing;Wang Hongyuan(School of Information and Mathematics,Changzhou University,Changzhou 213164,China)

机构地区：[1]常州大学信息数理学院,江苏常州213164

出　　处：《南京理工大学学报》2020年第5期614-623,共10页Journal of Nanjing University of Science and Technology

基　　金：国家自然科学基金(61806026,61976028,61572085,61502058);江苏省自然科学基金(BK20180956);江苏省教育科学“十三五”规划课题(B-a/2018/01/41)。

摘　　要：针对传统聚类算法在开始运行时要指定聚类数的问题,提出了一个贝叶斯可能性聚类模型,推理出了一种具有良好抗噪能力的可以估计聚类个数的可能性聚类方法。使用贝叶斯推理和粒子滤波推理为模型找出最大后验参数值,并利用泊松分布估计聚类数的最优值。在求解样本隶属度的过程中隶属度的值仅与样本点和其对应的聚类中心的距离有关,从而模型的抗噪能力得到提高。在UCI标准库数据集、真实医学图像数据集上对所提出的模型进行检验。结果表明,该文模型具有更好的聚类性能和抗噪能力。In order to solve the problem that the number of clusters must be specified when the traditional clustering algorithm starts to run,a Bayesian possibility clustering model is proposed,which can estimate the number of clusters and achieve satisfactory noise resistance ability.Bayesian inference and particle filter inference are used to find the maximum-a-posteriori parameters.Poisson distribution is used to estimate the optimal clustering number.In the process of computing the sample membership,the value of membership is only related to the distance between the sample point and its corresponding cluster center,which greatly improve the anti-noise ability of the model.Extensive experiments are performed on the data-sets of UCI standard library and the medical images collected in real-world environments.The experimental results show that the proposed method has better clustering performance and anti-noise ability.

关 键 词：模糊聚类 贝叶斯推理 粒子滤波推理 泊松分布 医学图像 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]