Seminar on Computational Learning and Adaptation



Minimum Message Length Applied to
Segmenting Simple Time Series


Jonathan J. Oliver
(joint work with Catherine Forbes)

Ultimode Systems
Data Mining Consultancy
Berkeley, CA
www.ultimode.com



Keywords: Segmentation, Minimum Message Length, MML, Bayes Factors, Evidence, Time Series

The segmentation problem arises in many applications in data mining, A.I. and statistics. In this talk, we consider segmenting simple time series -- this involves determining how many distinct intervals there are in a time series and when they occur. For example, when we examine ecomonic time series it would be useful to identify periods of growth, recession, depression, etc. We apply (without mathematical details) Minimum Message Length (MML) to the segmentation problem. We also consider a range of other approaches to segmentation, including: a Bayes Factors approach, Minimum Description Length (MDL) and Classical Statistical approaches. We find the segmentation problem interesting because it highlights significant differences between MML, MDL and Bayes Factors. Simulations comparing these approaches indicated that: a) MML gave significantly different and superior results to the Bayes Factors approach, and b) while MDL messages were shorter than MML messages, the MML results were again superior to MDL. Finally, we apply the segmentation method to real world time series data. 



Date: Thurs., February 12; Time: 4:15-5:30PM; Place: Gates 100

The goal of this seminar is to increase communication among local researchers with interests in computational approaches to learning and adaptation. If you would like to be added to (or removed from) the mailing list, or if you are interested in giving a talk in the seminar, please send email to iba@isle.org.


Return to seminar schedule.