JISE


  [ 1 ] [ 2 ] [ 3 ] [ 4 ] [ 5 ] [ 6 ] [ 7 ] [ 8 ] [ 9 ] [ 10 ] [ 11 ] [ 12 ] [ 13 ] [ 14 ] [ 15 ] [ 16 ]


Journal of Information Science and Engineering, Vol. 33 No. 2, pp. 517-536


DAPs: Mining using Change-Point Detection of Epileptic Activity Time Series Data


SUN-HEE KIM1, LEI LI2, CHRISTOS FALOUTSOS3, HYUNG-JEONG YANG4
AND SEONG-WHAN LEE
1
1
Department of Brain and Cognitive Engineering
Korea University
Seoul, 136-713 South Korea

2Computer Science Division
University of California
Berkeley CA, 94720 USA

3School of Computer Science
Carnegie Mellon University
Pittsburgh PA, 15213 USA

4Department of Computer Science
Chonnam National University
Gwangju, 500-757 South Korea
E-mail: sunheekim@korea.ac.kr;
lilei22@baidu.com; christos@cs.cmu.edu;
hjyang@jnu.ac.kr; swlee@image.korea.ac.kr


     The goal of this study is to mine meaningful patterns effectively and efficiently via change-point detection of the time series data, with the assistance of domain knowledge and observed data. With those patterns, our method can do segmentation and compression. We developed a novel gray-box approach for mining such data: Domain Assisted Parameter semi-free wave mining (DAPs). DAPs is intended for mining time series with rich domain-specific knowledge based on a chaos model. Specifically, it automatically detects a change-point of time sequences, respecting the minimal description length principle. And the time sequence is segmented based on the detected change-point, and each segment is fitted with a consistent model. The experimental results using both synthetic and real EEG data indicated that the developed method offers a significant improvement in segmentation and compression via pattern detection over other existing methods. DAPs reduced the number of bits of the observed data by detecting the changes in the patterns contained therein and brought about a higher average compression ratio, 1.6% more than WT (level 5). DAPs provides the advantages of (a) being capable of automatically detecting meaningful patterns, (b) being parameter semi-free, and (c) resulting in a huge reduction in data storage. These findings provide possible applications in the use of various medical devices that produce vast amounts of physiological data that should be monitored. 


Keywords: segmentation, gray-box model, chaos population model, parameter estimation, minimum description length, electroencephalography, compression

  Retrieve PDF document (JISE_201702_14.pdf)