El próximo viernes 23 de marzo a las 12:30h tendrá lugar, en laboratorio del grupo SIMD, el seminario Learning Probabilistic Decision Graph models from incomplete data, que será impartido por el Dr. Jens D. Nielsen.

Un resumen de la temática del mismo puede leerse a continuación:

This is a first study of how to apply the famous
framework of Nir Friedman (his key publications on this appeared in NIPS
97 and UAI 9 to the PDG model. We use a standard model selection
algorithm employing a decomposable score metric and local operators for
modifying a current model. However, in the presence of incomplete data
the score metric does not decompose as the sufficient statistics are not
necessarily available. The exact computation of the score metric would
require us to average over all possible completions of the incomplete
database, making exact computation of score differences intractable in
general. Instead of performing exact computations, we can approximate
the score metric by using instead of the sufficient statistics (that we
do not have) the expected sufficient statistics. We iterate between 1)
using the current model to compute expected score improvements, and 2)
modifying the current model to maximize the expected score improvement.
This leads to a standard Expectation-Maximization algorithm for
structural learning in the presence of incomplete data.