Improved piecewise vector quantized approximation based on normalized time subsequences

Abstract

Piecewise vector quantized approximation (PVQA) is a dimensionality reduction technique for time series data mining, which uses the closet codewords deriving from a codebook of key subsequences with equal length to represent the long time series. In this paper, we proposed an improved piecewise vector quantized approximation (IPVQA). In contrast to PVQA, IPVQA involves three stages, normalizing each time subsequence to remove the mean, executing the traditional piecewise vector quantized approximation and designing a novelly suitable distance function to measure the similarity of time series in the reduced space. The first stage deliberately neglects the vertical offsets in the target domain so that the ability of the codebook obtained from the training dataset is more powerful to represent the corresponding subsequences. The new function based on Euclidean distance in the last stage can effectively measure the similarity of time series. Experiments performing the clustering and classification on time series datasets demonstrate that the performance of the proposed method outperforms PVQA.