THE METHOD OF TIME GRANULARITY DETERMINATION ON TIME SERIES BASED ON STRUCTURAL SIMILARITY MEASURE ALGORITHM
Gao Xuedong, Hailan Chen · ISAHP proceedings · 2016
With the constant progress of science and technology, data size is increasing in the areas of social life and industrial production.And people are gradually aware of the potential value of data, causing a flood of big data and data mining.Data in real life is mostly related to time, called time series.At present, data analysis and data mining for time series has become the research focus in the field of data mining.Dimension reduction, similarity matching, cluster analysis and etc. for time series of large scales can dig out the useful information so that effectively guide the business decision and production application.According to the method of granular computing [1], information granularity affects the complexity and the validity of problems.In general, as the granularity decreases, the result becomes more accurate, while the computational complexity rises sharply.However it is likely to cover the nature of problems, so the size of granularity produces a great impact on the results of problems.Time granularity is the information granularity of time series, so time granularity is important for time series researches.In year 1996 to 1998, Claudio Bettini et al [2, 3] firstly proposed the concept that time interval can be regarded as time granularity.In year 2014, Xu Jianfeng et al [4] proposed the basic model on time granularity determination of multi-granulation time series, and used it to cluster analysis.But the model stays the level that time granularity is made by men, and there is no scientific calculation method is put forward.Granular computing theory [5] mainly transform time granularity according to the different problems, so as to simplify the problem solving.In this paper, the method for determining the optimal time granularity is different from traditional granular computing theory, it is based on basic granularity to determine the optimal time granularity.The wavelet analysis method is widely applied in the field of data mining, such as time series trend information extraction, time series similarity matching, time series data dimension reduction, etc. [6].The wavelet analysis method is firstly applied in time series pattern recognition in paper [7].Based on it, Zhang Haiqin et al [8] put forward the method of time series similarity pattern matching based on the wavelet transform.In year 1999, Zbigniew R et al [9] used Haar wavelet transform in time series similarity expression.The research of time granularity determination on time series is an important problem in data mining research.With the rapid growth of data quantity, it is necessary to obtain valuable information quickly and accurately by time series data mining.In order to solve the problem, in this paper, a method of time granularity on time series is proposed based on structural similarity measure algorithm.In this paper, the preliminary researches are carried on in the following three aspects.(1) Fluctuation point recognition Obviously there is an essential relationship between the size of time granularity and the fluctuation of time series, so the fluctuation point recognition of time series is particularly important.In this paper, it is proposed the definition of fluctuation point and its recognition method, we can extract the key information of a curve, and for further analysis.(2) Haar wavelet transform Conducted multi-scale decomposition of time series data by Haar wavelet transform, each decomposition time series data is decomposed into scale signal contains low frequency component and noise signal contains high frequency component and noise.Each Haar wavelet decomposition, the time granularity of Haar wavelet decomposition time series is i