Título: Structure recovery for partially observed discrete Markov random fields on graphs
Palestrante: Florencia Leonardi (IME-USP)
Data: 18/10/2021
Horário: 15:00hrs às 16:00hrs
Local: Transmissão online
Confira AQUI o link para a transmissão.
Resumo: Discrete Markov random fields on graphs, also known as graphical models in the statistical literature, have become popular in recent years due to their flexibility to capture conditional dependency relationships between variables. They have already been applied to many different problems in different fields such as Biology, Social Science, or Neuroscience. Graphical models are, in a sense, finite versions of general random fields or Gibbs distributions, classical models in stochastic processes. This talk will present the problem of estimating the interaction structure (conditional dependencies) between variables by a penalized pseudo-likelihood criterion. First, I will consider this criterion to estimate the interaction neighborhood of a single node, which will later be combined with the other estimated neighborhoods to obtain an estimator of the underlying graph. I will show some recent consistency results for the estimated neighborhood of a node and any finite sub-graph when the number of candidate nodes grows with the sample size. These results do not assume the usual positivity condition for the conditional probabilities of the model as it is usually assumed in the literature of Markov random fields. These results open new possibilities of extending these models to situations with sparsity, where many parameters of the model are null. I will also present some ongoing extensions of these results to processes satisfying mixing type conditions. This talk is based on a joint work with Iara Frondana and Rodrigo Carvalho and some work in progress with Magno Severino.
Todas os seminários são ministrados em inglês.
Os vídeos dos seminários passados estão disponíveis nos links abaixo:
Para o segundo semestre, alguns dias depois dos seminários, às gravações ficaram disponíveis AQUI.
