Research on Nearest Neighbor Classifying Method in Time Series Based on KPCA-CDTW

Conventional Dynamic Time Warping (DTW) warps the time axis (X-axis) only by considering the numerical similarity of time series, ignoring the offset on the numerical axis (Y-axis) and being sensitive to singular values, so the similarity of two time series cannot be accurately measured. Firstly; to solve "singularity" problem, this paper improves the local distance of traditional DTW, the new method is called CDTW (Combination Dynamic Time Warping), which combines the derivative feature with numerical feature of time series in the form of summation, the numerical feature K can be found by simple adjustment. Secondly; we integrate CDTW and KPCA (Kernel Principal Component Analysis) as KPCA-CDTW to reduce the dimensionality and the time complexity. Finally, we classify the time series by the nearest neighbor (1-NN) classification algorithm to test the validity of KPCA-CDTW, the experimental results on Coffee, Strawberry and other four datasets show that the new method has higher classification accuracy and much less classification time than other algorithms, which proves the effectiveness of this method.