Making the nearest neighbor meaningful for time series classification

The effectiveness of nearest neighbor search heavily relies on the definition of distance function. Unfortunately, the meaningfulness of the frequently used distance, such as Euclidean distance, fractional distance and so on, will degrade with the increasing dimensionality. This problem, which is called distance concentration or instability, makes NN method perform poorly in a proximity query. The most popular distance function for time series, dynamic time warping(DTW), also concentrates when it is used in high-dimensional space. We learn the exponent p of the norm based on nearest neighbor large margin criterion for time series classification. The distance concentration is countered by maximum discrimination instead of maximum variance of distance distribution. The empirical results we presented demonstrate that the proposed approach shows a uniformly behavior, with results comparable to classic 1NN-Euclidean and 1NN-DTW.