Efficient retrieval of similar time sequences under time warping

Fast similarity searching in large time sequence databases has typically used Euclidean distance as a dissimilarity metric. However, for several applications, including matching of voice, audio and medical signals (e.g., electrocardiograms), one is required to permit local accelerations and decelerations in the rate of sequences, leading to a popular, field tested dissimilarity metric called the "time warping" distance. From the indexing viewpoint, this metric presents two major challenges: (a) it does not lead to any natural indexable "features", and (b) comparing two sequences requires time quadratic in the sequence length. To address each problem, we propose to use: (a) a modification of the so called "FastMap", to map sequences into points, with little compromise of "recall" (typically zero); and (b) a fast linear test, to help us discard quickly many of the false alarms that FastMap will typically introduce. Using both ideas in cascade, our proposed method achieved up to an order of magnitude speed-up over sequential scanning on both real and synthetic datasets.