Presentation 2006-10-11
Bayesian Hypothesis Testing in Singular Models; a Case Study of Time Series Analysis
Kaori FUJIWARA, Sumio WATANABE,
PDF Download Page PDF download Page Link
Abstract(in Japanese) (See Japanese page)
Abstract(in English) Statistical hypothesis testing is constructed by comparing the probability of an alternative hypothesis with that of a null hypothesis. The log likelihood ratio (LR) using the maximum likelihood estimator MLE is not appropriate for testing singular models, because LR using MLE diverges in singular models resulting that it gives weak hypothesis testing. In this paper, based on algebraic geometrical method, we theoretically derive the asymptotic distribution of the Bayesian log likelihood, which show that the Bayesian testing hypothesis is appropriate for singular learning machines. The proposed method is applied to the problem of change point detection, which is a typical singular learning machine.
Keyword(in Japanese) (See Japanese page)
Keyword(in English) singular learning machine / Bayes hypothesis testing / Bayes marginal likelihood ratio / Bayes factor / change-point detection
Paper # NC2006-51
Date of Issue

Conference Information
Committee NC
Conference Date 2006/10/4(1days)
Place (in Japanese) (See Japanese page)
Place (in English)
Topics (in Japanese) (See Japanese page)
Topics (in English)
Chair
Vice Chair
Secretary
Assistant

Paper Information
Registration To Neurocomputing (NC)
Language JPN
Title (in Japanese) (See Japanese page)
Sub Title (in Japanese) (See Japanese page)
Title (in English) Bayesian Hypothesis Testing in Singular Models; a Case Study of Time Series Analysis
Sub Title (in English)
Keyword(1) singular learning machine
Keyword(2) Bayes hypothesis testing
Keyword(3) Bayes marginal likelihood ratio
Keyword(4) Bayes factor
Keyword(5) change-point detection
1st Author's Name Kaori FUJIWARA
1st Author's Affiliation Tokyo Research Lab., IBM Japan Ltd.()
2nd Author's Name Sumio WATANABE
2nd Author's Affiliation PI Lab., Tokyo Institute of Technology
Date 2006-10-11
Paper # NC2006-51
Volume (vol) vol.106
Number (no) 279
Page pp.pp.-
#Pages 5
Date of Issue