Locally Weighted DTW Method Based on Sparse Path Matrix

Dynamic time warping algorithm is one of the most effective time series similarity measures. The LDTW algorithm solves the problem that the dynamic time warping algorithm timing alignment method treats the instance features equally and ignores its local discriminative characteristics, but the time complexity of the LDTW generation of the weight matrix is too high. This paper proposes a LDTW algorithm based on sparse path matrices. In the process of iteratively generating a weight matrix from the training set, a sparse path matrix between homogeneous sequences and a sparse path matrix between different types of sequences are generated according to the type of sequence in the training set. Weight matrix over a sparse path matrix. The proposed algorithm is simulated and verified on the UCR dataset. The LDTW algorithm based on the sparse path matrix has the same classification error rate as LDTW, but the time cost of generating the weight matrix is significantly lower than the LDTW algorithm. It proves the effectiveness of path matrix LDTW in reducing time consumption.