Classificatory decomposition for time series clustering and categorization

• Authors: Jacek M. Zurada;Tomasz G. Smolinski
• Venue: Classificatory decomposition for time series clustering and categorization
• Year: 2004

The problem of time series categorization (or classification) is fundamentally concerned with the ability to classify time series into one of a determined number of predefined classes. The problem of time series clustering is a task in which one seeks to identify a finite set of categories (i.e. , clusters) to describe the temporal data. For these problems, algorithms extracting feature vectors based upon signal decomposition techniques have already been commonly used. Most of those algorithms have proven to be quite successful and useful in problems of time series clustering and classification. However, their main flaw lies in the fact that nowhere during the process of decomposition itself is the classification accuracy of the generated model taken into the consideration. Thus the reconstruction error can be low, but the ability to differentiate between classes, based on the decomposition, is not assured. This dissertation presents an attempt to extend the applicability and to improve the effectiveness of decomposition algorithms by providing them with “classification-awareness.” In this dissertation, a general framework for the methodology of classificatory decomposition for time series classification and clustering is proposed. The methodology combines the agility of rough sets in their accuracy and efficiency for data classification and clustering problems with signal decomposition techniques' ability to explain the hidden, underlying behavior of the modeled system. The methodology is based upon an application of a multi-objective evolutionary algorithm with fitness functions that deal with two considerations simultaneously: the first constituent of the function evaluation is based upon the reconstruction error, while the second utilizes various rough sets-based classification accuracy measures proposed in this dissertation. This produces a novel approach to the process of time series decomposition: classification itself is the “driving force” in the process of decomposition.