Dynamic time warping under pointwise shape context

Abstract

Similarity measures are of fundamental importance in time series data mining. Dynamic Time Warping (DTW) is a quite popular measure because it handles time distortions well. However, DTW has an inherent shortcoming in that DTW can lead to pathological alignments between time series where a single point maps onto a large subsection of another time series. To overcome this problem, we propose a novel variant of DTW named SC-DTW. SC-DTW employs shape context, a rich local shape descriptor, to replace the raw observed values considered by conventional DTW. The main novelties of SC-DTW are (1) it deeply explores both the numerical nature and shape nature of time series; and (2) neighborhood information for each point is taken into account. SC-DTW can generate a more feature-to-feature alignment between time series and thus serves as a robust similarity measure. We test the performance of SC-DTW on UCR time series datasets using the one nearest neighbor (1NN) classifier. Compared with other well-established methods, SC-DTW provides better accuracy on 24 of 34 datasets.