[Todos CMAT] Seminario de Probabilidad y Estadística, viernes 2/5

gillanes en cmat.edu.uy gillanes en cmat.edu.uy
Mie Abr 30 15:53:03 UYT 2014


Hola a todos!

El viernes que viene, 2 de mayo, se realizará el seminario de probabilidad 
y estadística a las 10hs en el salón de seminarios del piso 14.

El seminario de este viernes será especial, ya que tendremos dos 
expositores invitados. De 10:00 a 10:50 hablará Badih Ghattas (Institut de 
Mathématiques de Marseille). De 11:00 a 11:50 expondrá Catherine Aaron 
(Université Blaise-Pascal Clermont II).

Al final están los títulos de las charlas y los resúmenes.

Si alguien quiere estar en la lista de mails del seminario, por favor me 
avisa. Los esperamos a todos! Abrazo grande!

--------------------------------------------------

Badih Ghattas, Institut de Mathématiques de Marseille
Title : Bayesian Networks over Clusters.

Summary : Inferring genes networks using transcriptomic data is nowadays a 
big challenge in bioinformatics. Such network may be seen as a graph where 
nodes represent genes and vertices represent regulation (inhibition or 
activation). Different approaches aim to construct regulation networks from 
microarray data; logical networks, differential equations, and Bayesian 
networks, which are concerned by this talk. Bayesian networks are directed 
acyclic graphs modeling the joint distribution of the covariates (genes) by 
capturing the conditional dependencies between genes. Their inference using 
microarray data needs a structure optimization, the graphs structure, and 
estimating the parameters of the conditional distributions within each 
node. The structure optimization is known to be NP-hard, and several 
algorithmic issues has been suggested in the literature. We tackle this 
problem using dimension reduction for the data based on a simple idea. 
Cluster the genes, choose a representative for each cluster, and infer a 
Bayesian network using the representatives. Clustering is done in both 
continuous and discrete context. In the continuous context, classical 
methods are used, representatives are to chosen to be the nearest 
individuals to the mean individual. For discrete data, we have used Mutual 
Information as dissimilarity measure for clustering. We have tested our 
approach on both real and simulated data sets, and have shown its 
effectiveness. Next Step should be using such networks for supervised 
classification.


------------------------------------------------------------


Catherine Aaron, Université Blaise-Pascal Clermont II.

Summary:
In this talk we study the statistical properties of an estimator of the
support of the density based on
the notion of $\alpha-$shape which is a restriction of the Delaunay
complex where only ``small'' simplices are kept.
We also present two possible applications of such a support estimation.
One is aimed at reducing the bias
of the kernel density estimation when the density is bounded bellow by a
positive constant on its support.
The second one is a new depth definition that is suitable for non-convex
support


Más información sobre la lista de distribución Todos