Quantifying and localizing state uncertainty in hidden Markov models using conditional entropy profiles - Inria - Institut national de recherche en sciences et technologies du numérique Access content directly
Conference Papers Year : 2014

Quantifying and localizing state uncertainty in hidden Markov models using conditional entropy profiles

Abstract

Abstract. A family of graphical hidden Markov models that generalizes hidden Markov chain (HMC) and tree (HMT) models is introduced. It is shown that global uncertainty on the state process can be decomposed as a sum of conditional entropies that are interpreted as local contributions to global uncertainty. An efficient algorithm is derived to compute conditional entropy profiles in the case of HMC and HMT models. The relevance of these profiles and their complementarity with other state restoration algorithms for interpretation and diagnosis of hidden states is highlighted. It is also shown that classical smoothing profiles (posterior marginal probabilities of the states at each time, given the observations) cannot be related to global state uncertainty in the general case.
Fichier principal
Vignette du fichier
compstat2014_guedon.pdf (433.37 Ko) Télécharger le fichier
Origin : Files produced by the author(s)
Loading...

Dates and versions

hal-01058278 , version 1 (26-08-2014)

Identifiers

  • HAL Id : hal-01058278 , version 1
  • PRODINRA : 314098

Cite

Jean-Baptiste Durand, Yann Guédon. Quantifying and localizing state uncertainty in hidden Markov models using conditional entropy profiles. COMPSTAT 2014 - 21st International Conference on Computational Statistics, The International Association for Statistical Computing (IASC), Aug 2014, Genève, Switzerland. pp.213-221. ⟨hal-01058278⟩
640 View
352 Download

Share

Gmail Facebook Twitter LinkedIn More