Please use this identifier to cite or link to this item: http://hdl.handle.net/10397/86521
DC FieldValueLanguage
dc.contributorDepartment of Computing-
dc.creatorFu, Tak-chung-
dc.identifier.urihttps://theses.lib.polyu.edu.hk/handle/200/6549-
dc.language.isoEnglish-
dc.titleTime series pattern matching, discovery & segmentation for numeric-to-symbolic conversion-
dc.typeThesis-
dcterms.abstractRecently, the increasing use of temporal data has initiated various research and development attempts in the field of data mining. Time series are an important class of temporal data objects and they can be easily obtained from financial and scientific applications, e.g. daily temperatures, prices of mutual funds and stocks. They are in fact major sources of temporal databases and undoubtedly finding useful time series patterns are of primordial importance. While most of the research communities have concentrated on the forecasting issues, discovery of hidden behavior and relationship within a time series or among a set of time series has so far not yet been fully addressed. Unlike the transactional databases with discrete/symbolic items, time series data are characterized by their numerical, continuous nature. Hence, time series data are difficult to manipulate. But when they can be treated as segments instead of data points, interesting patterns can be discovered and it becomes an easy task to query, understand and mine them. So, it is suggested to break down the sequences into meaningful subsequences and represent them symbolically. We term this process as numeric-to-symbolic (N/S) conversion and consider it as one of the most important components in time series data mining systems. In this thesis, various algorithms for N/S conversion is proposed. They include: a flexible temporal pattern matching scheme which attempts to locate the perceptually important points in the data sequence for similarity computation is first proposed. As to human's behavior in identifying patterns from time series, the frequently used patterns are typically characterized by a few critical points and these points are perceptually important in human's identification process and should also be taken into accounts in the pattern matching process. The proposed scheme follows this idea by locating those perceptually important points and attractive results have been obtained. Based on that, methods for discovering frequently appearing patterns from time series are developed. The raw numerical data sequence of certain length will undergo a clustering process using the Kohonen's self-organizing maps through which similar data sequences or patterns are grouped together and represented by a pattern symbol. With the new time series pattern matching scheme and the pattern discovery algorithm introduced, we propose to address the time series segmentation problem in a more flexible way so as to facilitate dynamic N/S conversion, i.e., to segment the time series irregularly. This is achieved by an evolutionary segmentation algorithm which works with the pattern matching scheme to make the cutting decisions. Simulation results on the time series of the Hang Seng Index as well as different Hong Kong stocks show that the proposed models are effective and yet efficient.-
dcterms.accessRightsopen access-
dcterms.educationLevelM.Phil.-
dcterms.extentx, 89 leaves : ill. ; 30 cm.-
dcterms.issued2001-
dcterms.LCSHData mining-
dcterms.LCSHTime-series analysis-
dcterms.LCSHHong Kong Polytechnic University -- Dissertations-
Appears in Collections:Thesis
Show simple item record

Page views

46
Last Week
0
Last month
Citations as of Mar 24, 2024

Google ScholarTM

Check


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.