Hierarchical IB Algorithm For Multichannel Time Series

Time series is also known as dynamic series.It is consisted of a set of measurements over time.As a special type of time series,multichannel time series which contains several different types of collections of observations in different channels can hold more complicate information than single channel time series.It has been wildly used in different fields.For the reason that this sort of data bases is becoming more and more,and how to manage and analyse the large amount of multichannel time series becomes a big challenge.There are two key issues need to be solved in the task of mining multichannel time series data.One is that an effective feature extraction method to find discriminative features for multichannel time series is needed.It is because that the time series in multichannel is always long and the extracted features are self-contained and somehow similar.The other key issue is that an effective method which can take all the information in channels into account for multichannel time series data mining is badly needed.It is because the special structure of multichannel time series that every channel of it contains different information which cannot neglect.In order to solve the two issues,together with an improved bag-of-patterns(BoP)feature extraction method,this paper proposed a novel unsupervised multichannel time series categorization method named hierarchical Information bottleneck(hIB)algorithm,which has two layers.The transmission of the two layers of hIB enables the method take full use of all the information in different channels,which leads a desired result of clustering.In the first layer of hIB,a greedy agglomerative technique to find a hierarchical clustering tree in a bottom-up fashion is introduced to make it possible that the information in different channels can transmit and influence each other.In the second layer,hIB uses the output of the first layer to divide the multichannel time series,which is the final result of the clustering.The experiments comparing with the classical clustering methods and the new methods which can deal with multichannel time series shows that the proposed hIB algorithm gets the highestclustering accuracy and has some stability.The result of the experiments demonstrates that the proposed hIB algorithm in this thesis can effectively solve the two key issues in clustering multichannel time series which means that the algorithm has an advantage over clustering multichannel time series.