Pseudometrics for time series data classification using the nearest neighbor algorithm

Abstract

We propose that pseudometric, a subadditive distance measure, has sufficient properties to be a good structure to perform nearest neighbor pattern classification. There exist some theoretical results that asymptotically guarantee the classification accuracy of k-nearest neighbor when the sample size grows larger. These results hold true under the assumption that the distance measure is a metric. The results still hold for pseudometrics up to some technicality. Whether the results are valid for the non-subadditive distance measures is still left unanswered. Pseudometric is also practically appealing. Once we have a subadditive distance measure, the measure will have at least one significant advantage over the non-subadditive; one can directly plug such distance measure into systems which exploit the subadditivity to perform faster nearest neighbor search techniques. This work focuses on pseudometrics for time series. We propose two frameworks for studying and designing subadditive distance measures and a few examples of distance measures resulting from the frameworks. One framework is more general than the other and can be used to tailor distances from the other framework to gain better classification performance. Experimental results of nearest neighbor classification of the designed pseudometrics in comparison with well-known existing distance measures including Dynamic Time Warping showed that the designed distance measures are practical for time series classification.