Research On Time Series Classification Based On Piecewise Vector Quantized Approximation

A time series is a high-dimensional vector composed of numerous data points in a continuous time.The classification decision-making of time series is one of the most important studies of management science.Traditionally,decision accuracy mainly depends on the ability of decision-maker combined with objective factors.Only an experienced doctor could find the illness hide in some simple electrocardiogram.Therefore,it is very significant to find an objective way for a quantitative analysis of time series classification.Vector quantized(VQ)is a lossy compression algorithm based on block encoding which combines some scalar data together and quantizes the whole data in the vector space.We pay attention to both similarity measure and classification standard based on piecewise vector quantized approximation(PVQA).This thesis makes three improvements to PVQA by adopting the Mahalanobis distance,multiple codebooks and perceptually important points,respectively.The innovations of this thesis are mainly summarized as follows:This thesis proposes a piecewise vector quantized approximation based on the Mahalanobis distance(MPVQA).The traditional PVQA may be affected by feature dimensions when adopting the Euclidean distance as similarity measure.Thus,we introduce the Mahalanobis distance into consideration as similarity measure of reconstruction and error calculation for time series classification.Experimental results on UCR datasets show that this new method can improve the representation and classification performance of time series.This thesis proposes a multiple codebook piecewise vector quantized approximation(MCPVQA).The traditional PVQA creates just only one codebook for the whole training dataset,and ignores the influence of category information.To remedy this,MCPVQP is proposed,which can generate a codebook for each time series class.Experimental results on UCR datasets show that MCPVQA has a better performance compared to both PVQA,MPVQA and the combination of MCPVQA and Mahalanobis distance.This thesis proposes a multiple codebook important time subsequence approximation(MCITSA).The linear segmentation of PVQA may cause some loss of significant features,which would affect the final classification performance.Thus,MCITSA takes perceptually important point into consideration so that the useful features can be reserve to a large extent.Experimental results show that MCITSA has a better performance than PVQA and MCPVQA with the same total number of codewords.