Locating motifs in time-series data

Abstract

Finding motifs in time-series is proposed to make clustering of time-series subsequences meaningful, because most existing algorithms of clustering time-series subsequences are reported meaningless in recent studies. The existing motif finding algorithms emphasize the efficiency at the expense of quality, in terms of the number of time-series subse-quences in a motif and the total number of motifs found. In this paper, we formalize the problem as a continuous top-k motif balls problem in an m-dimensional space, and propose heuristic approaches that can significantly improve the quality of motifs with reasonable overhead, as shown in our experimental studies. © Springer-Verlag Berlin Heidelberg 2005.