The Research Of Early Classification On Time Series

Time series is a set of chronological data points in a narrow sense.Nevertheless,in general,any ordered value type data can be seen as time series to deal with.Time series data is existed in every aspect of social life,and as the time goes on,the data volume continue to rise,such as industry,web,network,traffic,medicine and so on.Time series classification refers to assign samples to pre-defined categories.Classification is of great use,such as medical diagnosis,disaster prediction process control,road traffic et al.However,earlier classification is favorable for decision in many domains.Early classification is that make prospective accurate results as earlier as possible.That is,early classification attempt to optimize the earliness of classification in the case of meeting minimum accuracy,instead of maximize accuracy.Time series early classification is of great importance in some time-sensitive domain,such as health informatics,disaster prediction,intrusion detection,stock market forecast et al.This paper takes time series as research object,geared to the specific characteristics of early classification,we have discussed the early classification of univariate time series and multivariate time series separately.The main research work is as follows:(1)The early classification research on univariate time series.We proposed an early classification based on PAA.As for time series,the length of series increases with time,so the length called as dimensionality is a challenge in reality of classification.Due to the speciality of early classification,the majority of dimensionality reduction algorithms can't be used in the practical application.Piecewise aggregate approximation can be applied in early classification,and is easy to operate.In this paper,we put forward early classification algorithm PAA_ECTS on the basis of ECTS,which use PAA to reduce the dimension of original time series data,and in low-dimensional space to early classify the time series,then conduct experiments on 42 data sets to compare with ECTS,ECDIRE,RelClass,EDSC.As for the experimental results,we take Wilkerson rank and test,which indicate that our method is superior to others.(2)Research on multivariate time series early classification.Based on center sequence,we came up with center-sequence_ECMTS,and on the basis of classifier ensemble,we raised ECMTSEn.To the multi-number variables of MTS,this paper we use center sequence to merge MTS into a center sequence i.e.univariate time series,in this way,we reduce the variable-based dimension,subsequently we improve center-sequence_ECMTS,combine PAA with center sequence,reduce the time-based dimension,then proposed MTSECP,compared with other MTS early classification algorithm,MTSECP greatly lowered the complexity.We use 6 data sets to verify our method.Because MTS have several variables,and different variable carry different characteristics,what's more,different variable measure different information.In this paper,we proposed a MTS early classification algorithm based on ensemble called ECMTSEn.ECMTSEn take full advantage of the contribution and sensitivity of each variable,finally obtain the ensemble of classification result from each variable,which take the relation among variables into account.Subsequently,we apply feature subset election into ensemble,proposed FSECMTSEn.What's more,we use 9 common MTS data sets to verify our method,and take Wilkerson rank and test,which show our methods can reach fine results.