Comparision performances of feature-based time series clustering techniques via simulation studies

Clustering analysis is one of the data mining techniques that aims dividing a data set into homogenous groups such that the distance between individuals in the same group is as minimum as possible, the distance between individuals in the different group is as maximum as possible. Time series clustering (TSC) can be considered application of any clustering algorithm to a data set consisted of numerous time series. In here, time series is data set consisted of observations measured at successive time points. TSC techniques are divided into three categories as i) raw-based, ii) model-based, and iii)feature-based. Raw-based TSC techniques are based on directly applying the clustering algorithms to time series and therefore, their computational cost is very high in large-scale time series. In order to overcome this disadvantage of raw-based TSC techniques, feature based and model based TSC techniques that are based on converting time series into lower dimension space are proposed. While model-based TSC techniques are based on predicting a model for each time series and using the model parameters in clustering, feature-based TSC techniques are based on applying feature extraction method to each time series and using features extracted in clustering. In this thesis, feature-based TSC methods are focused. So far, many feature-based TSC techniques have been developed. But, a comprehensive study has not been encountered aimed at comparison of performances of these methods. In this thesis, it is aimed determining of clustering algorithm and time series's features that provide best performance in TSC via simulation study. In the direction of this purpose, 14 clustering algorithms and 13 feature extraction methods have been compared by performing five simulation studies according to correctly clustering success. At the result of analyses, it is observed that Discrete Wavelet Transform and Ward linkage clustering algorithm have best performance. Keywords: Data mining, clustering analysis, time series clustering, feature-based time series clustering