Room: BBL 061
Speaker: Silja Renooij
Title: Sensitivity in Graphical Models: Bayesian Network Sensitivity Functions for Arc Removal and Hidden Markov Models
Bayesian networks are a member of the family of graphical models
that provide a graph-based representation for encoding a probability
distribution over a multi-dimensional space. A Bayesian network combines
a directed acyclic graph, representing the independences between
variables, with discrete conditional probability distributions for each variable.
A large number of probability parameters is required to completely
specify the conditional probability distributions, and these are
bound to be inaccurate. Sensitivity analysis is a general technique
to study the effects of changes in network parameters on some output
of interest. Our broad experience in sensitivity analysis for
Bayesian networks had led to several new insights and inventions.
In this presentation, I will demonstrate that analysing the effects
of removing an arc from the graph of a Bayesian network can also
be considered a type of sensitivity analysis. In addition, I will
show that results stemming from sensitivity research in Bayesian networks apply to another member of the graphical model family:
the Hidden Markov model; from these results we now have an efficient
algorithm for sensitivity analysis in such models as well."