Fast clustering for time-series data with average-time-sequence-vector generation based on dynamic time warping

This paper proposes a fast clustering method for time-series data based on average time sequence vector. A clustering procedure based on an exhaustive search method is time-consuming although its result typically exhibits high quality. BIRCH, which reduces the number of examples by data squashing based on a data structure CF (Clustering Feature) tree, represents an effective solution for such a method when the data set consists of numerical attributes only. For time-series data, however, a straightforward application of BIRCH based on a Euclidean distance for a pair of sequences, miserably fails since such a distance typically differs from human's perception. A dissimilarity measure based on DTW (Dynamic Time Warping) is desirable, but to the best of our knowledge no methods have been proposed for time-series data in the context of data squashing. In order to circumvent this problem, we propose DTWS (Dynamic Time Warping Squashed) tree, which employs a dissimilarity measure based on DTW, and compresses time sequences to the average time sequence vector. An average time sequence vector is obtained by a novel procedure which estimates correct shrinkage of a result of DTW. Experiments using the Australian sign language data demonstrate the superiority of the proposed method in terms of correctness of clustering, while its degradation of time efficiency is negligible.