Extending The Classification Paradigm To Multivariate Time Series

The problem of time series pattern recognition has drawn intensive attention from the data mining community, due to its prevalence in numerous applications.In most applications, especially with the wide-spread proliferation of multimedia technology, more and more multivariate time series data have appeared. These include motion capture data, bioinformatics sequence, financial data, moving objects tracking, human and computer interface and many others. However, most progress achieved in this area has focused on addressing univariate time series problems. Therefore, there is a strong demand for providing solutions that address these types of multivariate data for time series retrieval, especially for pattern recognition applications.In this thesis, we propose CMM, a simple yet effective Classification Model that addresses the characteristics of Multivariate Time Series Data. CMM, which is based on decision tree, has two parts: data preprocessing and decision tree classification. Under CMM, data preprocessing consists of two phases, namely feature extraction and dimension ranking and selection. In the feature extraction phase, rather than treat each dimension of the multivariate data separately as other methods do, C3M considers all the dimensions altogether when segmentation is performed to the sequence. Then for each segment, C3M extracts only one coefficient of the Chebyshev polynomials that are used to approximate that segment as the local feature. Dimension ranking and selection adopts a supervised manner, in which feature dimensions are ranked according to their combined information gain ratio, an indicator of their predictive ability. C3M then selects the Top-K predictive dimensions, and features in those K dimensions are used to vectorize the original multivariate data. These feature vectors obtained in data preprocessing phase serve as the input to the Decision Tree classifier in the subsequent classification part. Experimental results show that C3M achieves significant performance improvements in terms of both classification accuracy and processing time compared to conventional schema...