On the Global Features and Nuclear Field Clustering of Long Time Series
首发时间:2014-07-30
Abstract:Cluster analysis is an important and challenging subject in time series data mining. It has a very important application prospects in many areas, such as medical images, atmosphere, finance, etc. Many current clustering techniques still have many problems, for example, k-means is a very effective method in finding different shapes and tolerating noise, but its result severely depends on the suitable choice of parameters. Inspired by nuclear field in physics, we propose a new dynamic clustering method based on nuclear field and global features for long time series. Basically, each time series is mapped to one data point in global features space, then the data point is considered as a material particle with a spherically symmetric field around it and the interaction of all data points forms a nuclear field. Through the interaction of nuclear force, the initial clusters are iteratively merged and a hierarchy of clusters is generated. Experimental results show that compared with the typical clustering method k-means, the proposed approach enjoys favorite clustering quality and requires no careful parameters tuning.
keywords: Dynamic Clustering Data Mining Global Features Nuclear Field Field Force
点击查看论文中文信息
基于全局特征和核力场的长时间序列聚类
摘要:聚类分析在时间序列数据挖掘中是一项重要且有挑战性的课题,它在很多方面有非常重要的应用,如医学图像、气象、金融等等。许多已有的聚类技术仍然存在许多问题,如k均值是一种非常有效地寻找不同形状和容忍噪声的方法,但是它的结果极大程度上依赖于合适的变量选择。受物理中核力场的启发,本文提出一种基于长时间序列核力场和全局特征的新的动态聚类方法。每个时间序列被映射成全局特征空间的一个数据点,然后数据点被视为一个周围有球形对称场的质点,并且所有数据点之间的关系形成了一个核力场。通过核力的作用,初始的团被迭代合并进而形成一个簇的层次结构。实验结果说明相比传统的k均值聚类方法,本文提出的方法有更好的聚类质量并且不需要任何仔细的参数调整。
论文图表:
引用
No.4604629113314056****
同行评议
勘误表
基于全局特征和核力场的长时间序列聚类
评论
全部评论0/1000