Punt, H. (2000) Estimating Quality of Segmentation for Multistationary Time Series Modeling. Master's Thesis / Essay, Computing Science.
|
Text
Infor_Ma_2000_HPunt.CV.pdf - Published Version Download (1MB) | Preview |
Abstract
The main problem associated with time series modeling is that of modeling non—stationary time series. Monolitic, global models assume stationarity, e.g. the properties of the time series do not change with time. These models fail to capture the non—stationary dynamics of many real—world time series. One class of non—stationary time series are those that switch between multiple stationary regions at certain points in time. These multistationary time series series can be modeled by modular models which combine a segmentation of time series into stationary regions with a number of local models specialized in modeling these regions. The outputs of the local models are combined to form a single output. Assuming good local models can be found for stationary local data and that they can be combined successfully, the remaining problem is segmenting the stationary regions. During the construction of a modular model the quality of segmentation must therefore be evaluated, e.g. the stationarity of local data must be quantified. This is a problem because time series modeling does not make assumptions about the properties of the stationary regions, and no non—parametric estimate for stationarity exists. One solution is to assume that stationarity of local data is expressed by the the local models, estimating the quality of segmentation indirectly from their performance. This is the approach taken by modular models found in literature such as Gated Experts Networks. A number of problems are associated with this approach, caused by the interdependence between segmentation and local modeling. The goal of this thesis is was to find out if stationarity of local data can be expressed without constructing local models, with the aim of breaking the interdependence between segmentation and modeling. Based on a hypothesis that stationarity may be expressed by means of a number of traditional statistics, measures were developed to estimate segmentation quality directly from the statistical properties of local data. The viability of these measures of segmentation quality was subsequently tested in a number of experiments. Based on the results of these experiments, the main conclusion must be that it is not possible to generally estimate segmentation quality by means of estimating stationarity of local data using traditional statistics. The newly developed measures of segmentation quality can therefore not be used to break the interdependence between segmentation and local modeling in modular time series models.
| Item Type: | Thesis (Master's Thesis / Essay) |
|---|---|
| Degree programme: | Computing Science |
| Thesis type: | Master's Thesis / Essay |
| Language: | English |
| Date Deposited: | 15 Feb 2018 07:29 |
| Last Modified: | 15 Feb 2018 07:29 |
| URI: | https://fse.studenttheses.ub.rug.nl/id/eprint/8830 |
Actions (login required)
 |
View Item |