Skip to Main content Skip to Navigation

Réseau bayésien dynamique étiqueté : cadre et apprentissage de structure pour application aux réseaux écologiques

Abstract : An ecological network represents the interactions between living species within an ecosystem. The knowledge of the structure of such a network is an important challenge in the field of ecology.This task can be realized by inference methods : a set of methods that uses ecological observations data (species abundance, presence or absence of species...) in order to learn the interactions mathematically, by the exploitation of the effect of these interactions on the observed data.This thesis describes a case where the ecological data we dispose of are only data of presence/absence of species observed at different moments. The goal is to develop a method that exploits those kind of data in order to learn the interaction between these species. The main difficulty is that binary variables carry little information. Expert knowledge on the system is used to help learning the network's structure.We use the framework of dynamic Bayesian network : temporal presence/absence data are modeled as the realization of a set of dynamic random variables whose dependencies are described by an oriented graph. Such a model can be simplified using expert knowledge.This thesis describes a particular model of "labelled" dynamic Bayesian network. In this model, the graph is defined by a small number of different types of interactions that constitute a set of labels attributed to the edges of the graph.This model can describe several phenomena where an information or a perturbation can be propagated by contact (rumour, disease, forest fire....)This model describes the presence or absence probabilities of each species as a function of the number of interactions of each label this species is subject to. This model allows to describe every presence/absence probability of species using a small number of parameters independent from the network's structure. This is the framework used for the modeling of species dynamics within an ecological network : the information propagated is the presence or the absence of a species, knowing the interaction between the species of the network. Then, we describe the processes we use for learning the structure of a labelled dynamic Bayesian network using time series of binary variables. This 'Estimation-Restoration' algorithm alternates two steps : a phase of parameter estimation knowing the structure, and a phase of structure learning knowing the parameters. This last step can be complex. It is done by solving a integer linear programming problem. This allows to use efficient existing tools for solving those kind of problems. Moreover, we can easily add expert knowledge by the form of linear constraints. This process has been used on a particular case study :the observation of arthropods species trapped in experimental fields in the united kingdom. In order to highlight the differences between the different crops, different networks have been learnt. Finally, we compare the learnt network with others, learnt with different learning methods on the same data.
Document type :
Complete list of metadatas

Cited literature [130 references]  Display  Hide  Download
Contributor : Abes Star :  Contact
Submitted on : Wednesday, February 19, 2020 - 12:38:07 PM
Last modification on : Monday, May 25, 2020 - 1:52:07 PM
Long-term archiving on: : Wednesday, May 20, 2020 - 2:55:32 PM


Version validated by the jury (STAR)


  • HAL Id : tel-02484410, version 1



Etienne Auclair. Réseau bayésien dynamique étiqueté : cadre et apprentissage de structure pour application aux réseaux écologiques. Systèmes dynamiques [math.DS]. Université Paul Sabatier - Toulouse III, 2019. Français. ⟨NNT : 2019TOU30002⟩. ⟨tel-02484410⟩



Record views


Files downloads