Similarity measures for time series data classification using grid representation and matrix distance

摘要

Two similarity measures are proposed that can successfully capture both the numerical and point distribution characteristics of time series. More specifically, a novel grid representation for time series is first presented, with which a time series is segmented and compiled into a matrix format. Based on the proposed grid representation, two matrix matching algorithms, matrix-based Euclidean distance (GMED) and matrix-based dynamic time warping (GMDTW), are adapted to measure the similarity of matrix-like time series. Last, to assess the effectiveness of the proposed similarity measures, 1NN classification and K-means experiments are conducted using 22 online datasets from the UCR time series datasets Web site. In general, the results indicate that GMDTW measure is apparently superior to most current measures in accuracy, while the GMED can achieve much higher efficiency than dynamic time warping algorithm with equivalent performance. Furthermore, effects of the parameters in the proposed measures are analyzed and a way to determine the values of the parameters has been given.