Gu, Zhuoer (2017) Mining previously unknown patterns in time series data. PhD thesis, University of Warwick.

Preview

PDF WRAP_Theses_Gu_2017.pdf - Submitted Version - Requires a PDF viewer. Download (3600Kb) | Preview

Request Changes to record.

Abstract

The emerging importance of distributed computing systems raises the needs of gaining a better understanding of system performance. As a major indicator of system performance, analysing CPU host load helps evaluate system performance in many ways. Discovering similar patterns in CPU host load is very useful since many applications rely on the pattern mined from the CPU host load, such as pattern-based prediction, classification and relative rule mining of CPU host load.

Essentially, the problem of mining patterns in CPU host load is mining the time series data. Due to the complexity of the problem, many traditional mining techniques for time series data are not suitable anymore. Comparing to mining known patterns in time series, mining unknown patterns is a much more challenging task. In this thesis, we investigate the major difficulties of the problem and develop the techniques for mining unknown patterns by extending the traditional techniques of mining the known patterns.

In this thesis, we develop two different CPU host load discovery methods: the segment-based method and the reduction-based method to optimize the pattern discovery process. The segment-based method works by extracting segment features while the reduction-based method works by reducing the size of raw data. The segment-based pattern discovery method maps the CPU host load segments to a 5-dimension space, then applies the DBSCAN clustering method to discover similar segments. The reduction-based method reduces the dimensionality and numerosity of the CPU host load to reduce the search space. A cascade method is proposed to support accurate pattern mining while maintaining efficiency.

The investigations into the CPU host load data inspired us to further develop a pattern mining algorithm for general time series data. The method filters out the unlikely starting positions for reoccurring patterns at the early stage and then iteratively locates all best-matching patterns. The results obtained by our method do not contain any meaningless patterns, which has been a different problematic issue for a long time. Comparing to the state of art techniques, our method is more efficient and effective in most scenarios.

Item Type:	Thesis (PhD)
Subjects:	Q Science > QA Mathematics > QA76 Electronic computers. Computer science. Computer software
Library of Congress Subject Headings (LCSH):	Data mining, Pattern recognition systems, Time-series analysis -- Data processing, Computer systems -- Evaluation
Official Date:	September 2017
Dates:	Date Event September 2017 Submitted
Institution:	University of Warwick
Theses Department:	Department of Computer Science
Thesis Type:	PhD
Publication Status:	Unpublished
Supervisor(s)/Advisor:	He, Ligang
Sponsors:	China Scholarship Council
Format of File:	pdf
Extent:	xvi, 133 leaves : illustrations, charts
Language:	eng

Request changes or add full text files to a record

Repository staff actions (login required)

View Item

Downloads