Intership (M2) on ordinal regression models for transition matrix in HMM

31320 Castanet-Tolosan

Retour à la liste des résultats

Présentation INRAE

The National Research Institute for Agriculture, Food and Environment (INRAE) is a public research institution bringing together a working community of 12,000 people, with 272 research, service and experimental units located across 18 centers throughout France. INRAE ranks among the very top world leaders in agricultural and food sciences, as well as plant and animal sciences. Its research aims to build solutions for high-performing agriculture, quality food, and the sustainable management of resources and ecosystems.

Environnement de travail, missions et activités

You will be welcomed within the Mathematics and Informatics unit of the INRAE Occitanie-Toulouse center (MIAT, https://miat.inrae.fr/).

Context/Motivation. Hidden Markov Models (HMMs) have been extensively used to analyse time series whose dynamics can be described in terms of a hidden process evolving through successive states. Applications span a wide range of fields, including ecology, as well as medicine, speech analysis, and psychology. An HMM is defined by two sets of random variables: the hidden (latent) process which is a Markov chain, and the observations which are independent given the hidden chain and are modeled via an emission distribution. Hidden states are often discrete, representing regimes, classes, …

When the number of hidden states exceeds two or three, the transition matrix of the Markov chain rapidly becomes difficult to interpret and estimate, motivating more parsimonious structures, i.e. a parametric version. Existing works rely on regression models but are only ad-hoc without generality. 

Objective of the internship. Many applications naturally involve ordered states. Examples include low, medium and high abundance levels of a continuous ecological variable (e.g. soil seed-bank levels), stages of disease progression, or increasing activity levels ranging from sleep-like behaviour to low, medium and high activity.  The proposed project focuses on this category of hidden states and the objective is to propose a sound and parsimonious modelling approach for the transition matrix with ordinal hidden states.

Work plan. Weiß and Swidan (2025) proposed ordinal regression to model an ordinal observed variable in an HMM. More specifically, the observed variable is linked to a continuous latent variable emitted conditionally on the hidden state through threshold parameters, so that the emission distribution can, for instance, be formulated as an ordinal logit model.

Our objective is to extend this approach to the modelling of the transition matrix for ordinal hidden states. The first step is to formalise the corresponding HMM. Several options are possible and will be explored. For instance, emissions could follow the same ordering than hidden states or not, leading to different modelling complexity. A key question will be to accommodate the strong persistence typical to HMM, which may violate the assumption of proportional odds in ordinal regression. Then we will address the question of parameter estimation via the classical algorithm for HMM inference: EM and the forwad-backward algorithm. The second step will be to develop the associated EM algorithm and to evaluate its behaviour. More specifically, the following steps are planned:

  • bibliography on HMM, EM, ordinal/cumulative logistic models 
  • writing the model and coding a sampler from this model
  • writing EM equations, in particular E step
  • coding EM
  • Numerical experiments on EM behaviour, using simulations

Depending on the progress of the project, extensions to Hidden Semi-Markov Models or Coupled HMMs or to more general models than ordinal logistic regression (with less restrictive assumptions) are possible. 

Weiß, H. C., & Swidan, O. (2025). Hidden‑Markov models for ordinal time series. AStA Advances in Statistical Analysis, 109, 217‑239. https://doi.org/10.1007/s10182-024-00514-1

Formations et compétences recherchées

Master/Ingénieur (Bac+5)

Competences needed. training in statistics; sufficient English level for scientific exchanges; interest in programming, especially in R. 

Votre qualité de vie à INRAE

En rejoignant INRAE, vous bénéficiez (selon le type de contrat et sa durée) :

- jusqu'à 30 jours de congés + 15 RTT par an (pour un temps plein)
- de dispositifs de développement des compétences : formation, conseil en orientation professionnelle ;
- de la prise en charge à 50 % de votre cotisation à la complémentaire santé collective ;
- d'un soutien à la parentalité : CESU garde d'enfants, prestations pour les loisirs ;
- d'un accompagnement social : conseil et écoute, aides et prêts sociaux ;
- de prestations vacances et loisirs : chèque-vacances, hébergements à tarif préférentiel ;
- d'activités sportives et culturelles ;
- d'une restauration collective.

Modalités pour postuler

Contact.  Nathalie Peyrard and Sandra Plancade at MIAT-INRAE (nathalie.peyrard/sandra.plancade@inrae.fr), and Sina Mews at Georg-August-University of Göttingen (sina.mews@uni-goettingen.de). To apply, please send your CV, cover letter, and undergraduate and graduate transcripts to the three email addresses (in English).

Les personnes accueillies à INRAE, établissement public de recherche, sont soumises aux dispositions du Code de la fonction publique notamment en ce qui concerne l’obligation de neutralité et le respect du principe de laïcité. A ce titre, dans l’exercice de leurs fonctions, qu’elles soient ou non au contact du public, elles ne doivent pas manifester leurs convictions, par leur comportement ou leur tenue, qu’elles soient religieuses, philosophiques ou politiques. > En savoir plus : site fonction publique.gouv.fr

Référence de l'offre

  • Contrat : Stage
  • Durée : 5 months from March to July 2027
  • Début du contrat : 01/03/2027
  • Rémunération : €4.50/month
  • N° de l'offre : OT-31042
  • Date limite : 16/12/2026

Contact

Venir en France

Notre guide des accueils internationaux