The comparison of performances of feature based time series classification method

Time Series Classification (TSC) is a special type of classification and can be defined as a process of predicting "class" of time series according to their characteristics properties. The methods used for this objective can be grouped under collected three main titles as i) distance based, ii) model-based, and iii)feature based. Distance-based TSC methods work with raw data and are based on determining similar time series by using a distance measure. The first step of model-based TSC methods is that a model which represents the stochastic behavior of the time series is predicted for each time series by using modeling techniques such as, Markov models and Hidden Markov models. The next step is based on using of predicted model structure instead of time series at the classification stage. Finally, in feature-based TSC methods, time series are transformed into feature space with lower dimension and these features instead of time series are used at the classification. The main subject of this thesis is feature-based TCS methods and to compare performances of the most known classification methods and of features of time series sets. In this context, 13 classification methods, 10 features of time series and 39 UCR data that are at the different categories are used. For comparisons, accuracy rate is calculated for each method and each data set consisting of features of time series and performance evaluation is carried out according to results obtained. Keywords: Classification, Time series classification, feature-based time series classification