Time Series Classification Based on Multi-codebook Piecewise Vector Quantized Approximation

Piecewise vector quantized approximation (PVQA) is a dimensionality reduction technique for time series data mining, which adopts the closet codeword stemming from a codebook of time subsequences with equal length to represent the long time series. This paper proposes a multi-codebook piecewise vector quantized approximation (MCPVQA), in which we generate a codebook for each class using PVQA on considering the difference between categories. Thus, each codebook contains the corresponding class information. In addition, training time series do not need to reconstruct for classification tasks. MCPVQA needs to only reconstruct an unseen time series using these codebooks and predict its class label according to reconstruction errors in each class. Experimental results on three time series datasets demonstrate that MCPVQA is more powerful to represent time series and has better classification performance than PVQA.