基于概率和代表点的数据流动态聚类算法

发布时间：2018-02-24 15:20

本文关键词： 数据流能量函数概率优化算法动态聚类　出处：《计算机研究与发展》2016年05期 　论文类型：期刊论文

【摘要】：为了解决数据流动态聚类问题,提出了一种概率化的基于代表点聚类算法.首先,基于概率框架给出了AP(affinity propagation)聚类算法和EEM(enhancedα-expansion move)聚类算法的联合目标函数,提出了概率化的基于代表点聚类算法;其次,根据样本与其代表点之间的概率,提出了基于概率的漂移动态α-expansion数据流聚类算法.该算法使得新数据的代表点尽可能贴近原始数据的代表点,从而提高聚类性能;另一方面,考虑到原始数据与新数据的相似性,该算法能够处理2种漂移过程中的动态聚类问题:1)新数据与原始数据分享部分数据,其余数据与原始数据相似;2)没有相同的数据,新数据与原始数据有相似关系.在人工合成数据集D31,Birch3以及真实数据集Forest Covertpye,KDD CUP99的实验结果均显示出了所提之算法能够处理数据流聚类问题,并保证聚类性能稳定.
[Abstract]:In order to solve the dynamic clustering problem of data flow, a probabilistic representative point clustering algorithm is proposed. Firstly, the joint objective functions of AP(affinity propagation clustering algorithm and EEM(enhanced 伪 -expansion clustering algorithm are given based on the probability framework. A probabilistic clustering algorithm based on representative points is proposed. Secondly, according to the probability between the samples and their representative points, In this paper, a probabilistic drift dynamic 伪 -expansion data stream clustering algorithm is proposed, which makes the representative points of the new data as close as possible to the representative points of the original data, thus improving the clustering performance. On the other hand, considering the similarity between the original data and the new data, This algorithm can deal with the dynamic clustering problem of two drift processes: 1) the new data share part of the data with the original data, and the other data are similar to the original data and do not have the same data. The experimental results of the synthetic data set D31 / Birch3 and the real data set Forest overtpyeh CUP99 show that the proposed algorithm can deal with the data flow clustering problem and ensure the clustering performance is stable.
【作者单位】：江南大学数字媒体学院;齐鲁工业大学信息学院;
【基金】：国家自然科学基金项目(6127220) 山东省高等学校科技计划项目(J14LN05) 江苏省普通高校研究生科研创新计划基金项目(KYLX_1124)~~
【分类号】：TP311.13

【相似文献】