摘要
针对可能性聚类算法对初始化参数敏感及容易产生重合聚类的问题,提出了多尺度可能性聚类算法(MPCM)。算法结合均值漂移聚类算法与可能性聚类算法的思想,使其既保留了均值漂移聚类算法中能够揭示数据的多尺度聚类结构、不依赖于初始化参数的优点,也保留了可能性聚类算法可对数据集进行模糊划分的优点。同时,避免了均值漂移算法计算量过大以及可能性聚类对容易产生重合聚类的缺点。与传统的可能性聚类及其改进算法的对比实验结果表明,MPCM能够更加准确地揭示数据在不同尺度下的聚类结构,具有相对较好的聚类性能。
To overcome the problems of generating coincident clusters and initialization sensitivity of the possibilistic clustering algorithm(PCM),a new clustering algorithm called multi-scale possibilistic clustering algorithm(MPCM) was proposed in this paper.MPCM is inspired by the possibilisty clustering algorithm and the mean shift clustering algorithm(MSC),which makes it inherit the merit of both.The PCM can give a fuzzy partition of the data set,and the MSC can indicate the cluster structure in different scales and is independent to the initializations.In the meanwhile,MPCM avoids the problems of both the MSC that it has a high computation and the PCM that it tends to generate coincident clusters.The contrast experimental results show that MPCM can indicate the data structure in different scales more accurately and have a relatively better performance.
出处
《长春理工大学学报(自然科学版)》
2010年第4期124-127,共4页
Journal of Changchun University of Science and Technology(Natural Science Edition)
基金
吉林省科技发展计划项目(20080353)
关键词
模糊聚类
可能性聚类
均值漂移
多尺度结构
fuzzy clustering
possibilistic clustering
mean shift
multi-scale structure