Research On Curve Registration And Similarity Measures For Time Series

There are a large variety of data generated all the time in the era of information explosion,so data mining technology need to be developed with the increase in data amount and data types.Time series data are common in data mining,for example:stock price,consumer price index,the local monthly mean temperature and humidity,electrocardiogram etc.In addition,some kinds of data can't be analyzed by existing data analysis techniques due to their own complexity,and then they are often converted into time series.For example,in the study of leaf shape,distance data,which measure from the edge of the leaf to the geometric center of the leaf,are often grouped into time series according to the adjacent position relation.And then some traditional time series data analysis methods can be used to analyze them.Time series can be modeled easily and related technology is more mature,so it is easier to analyze time series directly and more value information can be obtained from them.The main research contents of time series data mining include prediction modeling?clustering?classification,etc.Similarity measure between data objects is an important prerequisite for studying above problems.The results of similarity measure methods will seriously affect the results of data mining,so how to improve the correlation between data objects is an important task for time series data analysis.We are likely to encounter a class of data that has a large similarity in shape and trend but with time warping(time difference)which will seriously affect the similarity measurement results and is easy to produce misclassification.In addition,how to extract attributes and define time series through attributes will also affect the relevance of the data analysis results.Therefore,in response to curve registration and similarity measurement methods between time series,the main works in this thesis includes the following two aspects:(1)Improving the efficiency and the effect of curve registration.Two kinds of non-uniform sampling method are proposed to align the curve: slope based non-uniform sampling(SBNS)and arc length based non-uniform sampling(ALBNS).The former method samples according to the slope size of the function curve,while the later method samples evenly in the arc length of function curve.The proposed two kinds of non-uniform sampling approaches sample according to characteristics of curve instead of sampling evenly in the time axis,which can overcome the limits of uniform sampling method that distributes the number and location of sample points improperly to improve the effect of curve registration.The experimental results on simulated data and real data show that the proposed two kinds of non-uniform sampling approaches are better than uniform sampling in time efficiency and the effect of curve registration.(2)Improving the clustering accuracy and efficiency.Two time series data similarity measure methods are put forward:The maximum shifting correlation coefficient similarity measure method(MSCD),and the attribute transformation based similarity measure strategy(AT).The former can avoid bad clustering results caused by time warping(phase difference)through registration of any two time series data.The later can improve the effect of time series clustering by attributes transformation of time series.Experimental results show that the proposed two similarity measures can improve the efficiency and accuracy of time series clustering.This thesis studies the curve registration and similarity analysis of time series data deeply.The proposed methods about these two aspects may lay the foundation for the analysis of time series data such as clustering? classification and forecasting modeling,etc.