针对时间序列数据挖掘的双加权聚类集成  

Bi-weighted clustering ensemble toward time series data mining

作　　者：胡健 王海林 肖鹏 尹君 Hu jian;Wang Hailin;Xiao peng;Yin jun(Yunnan power grid information center,Yunnan,Kunming 650217)

机构地区：[1]云南电网有限责任公司信息中心,昆明650217

出　　处：《云南电力技术》2021年第1期76-81,86,共7页Yunnan Electric Power

摘　　要：静态数据的聚类方法已得到了较为深入的研究,然而现实生活中越来越多的应用涉及到时间序列的聚类分析。但此类数据具有复杂的动态特性、高维度和海量性等特点,使得传统的算法无法获得较为理想的聚类结果。本项目将针对时间序列数据挖掘的实际需要,利用集成学习的最新研究成果,深入研究基于生成模型和特征表示的时间序列数据挖掘技术,并提出了两种时间序列聚类集成模型,以解决以下主要问题:(1)通过提出新的聚类集成模型,解决时间序列聚类算法中的初始化敏感。(2)通过引入双加权机制,使得聚类集成学习的可靠性及性能得到进一步的提高。(3)在时间序列聚类分析中能够有效地捕捉类簇的本征结构,能自动识别类数。(4)提高聚类集成模型的延展性,能根据不同特点的时间序列数据,选择不同的模型设置和多特征表示。综上所述本课题将在集成学习算法的框架下,对时间序列数据聚类分析提出较为前沿的理论研究,其研究成果必将具有较高的理论和实用价值。Although conventional static data clustering has been studied for many years,time series clustering recently has been becoming quite popular in various fields due to that such underpinning techniques can discover the intrinsic structures and condense or summarize information contained in growing time series datasets.Unlike static data,time series have many distinct characteristics,including high dimensionality,complex time dependency,and large volume,all of which make the clustering of time series more challenging than static data clustering.In this proposal,we intensively study the time series clustering algorithms from generative model-based and representation-based methodologies respectively,and proposed two ensemble learning approaches for time series clustering problems.Four key issues are explored in this proposal:(1)the clustering ensemble technique is used to tackle model initialization problems.(2)A novel Bi-weighting scheme is proposed to optimally reconcile the input partitions into a single consolidated clustering solution,where both the partition and the clustering weights are intrinsically derived from the objective function of clustering ensemble without any prior information.(3)The proposed approaches have out-standing ability in automatic detection of cluster number.(4)The proposed approaches have enhanced flexibility in association of different model setting and feature representations.Sum of all,we will carry out the forefront research of time series clustering in association of clustering ensemble techniques,the research results will not only contribute to the theoretical analysis,but also applications of time series data mining and pattern recognition.

关 键 词：聚类 集成学习 非监督学习 数据挖掘 时间序列 

分 类 号：TM74[电气工程—电力系统及自动化]