Fast Similarity Search in the Presence of Noise, Scaling, and Translation in TimeSeries Da.pdfVIP

Fast Similarity Search in the Presence of Noise, Scaling, and Translation in Time-Series Databases Rakesh Agrawal King-Ip Lin Harpreet S. Sa whney Kyuseok Shim IBM Almaden Research Center 650 Harry Road, San Jose, CA 95120 Abstract  Identify companies with similar pattern of growth. We intro duce a new mo del of similarity of time se-  Determine pro ducts with similar selling pat- quences that captures the intuitive notion that two terns. sequences should b e considered similar if they have enough non-overlappi ng time-ordered pairs of subse-  Discover sto cks with similar price movements. quences thar are similar. The mo del allows the am- plitude of one of the two sequences to b e scaled by any  Find p ortions of seismic waves that are not suitable amount and its oset adjusted appropriately. similar to sp ot geological irregularities. Two subsequences are considered similar if one can b e enclosed within an envelop e of a sp ecied width drawn In [1], an indexing structure w as prop osed for around the other. The mo del also allows non-matching fast similarity searches over time-series databases, gaps in the matching subsequences. The matching sub- assuming that the data as well as query sequences sequences need not b e aligned along the time axis. were of the same length. They use the Discrete Given this mo del of similarity, we present fast search Fourier Transform (DFT) to map a time sequence techniques for d