Learning inference friendly Bayesian networks: using incremental compilation
Studenteropgave: Speciale (inkl. HD afgangsprojekt)
- Søren Pedersen
- Martin Karlsen
2. semester, Datalogi, Kandidat (Kandidatuddannelse)
This report describes a project with the aim
of exploring structural learning of Bayesian
networks. Specifically the complexity of the
generated network, as a result of the chosen
learning method.
We examine the junction tree method for do-
ing propagation in Bayesian networks, de-
scribing the steps in compiling the junction
tree from the network structure. From this
analysis we learn that one cause of complex-
ity in junction tree is the size of the cliques.
A score function which scored a network di-
rectly on the combined size of the cliques and
the log-likelihood are proposed. This func-
tion uses a parameter to weight whether the
complexity versus the likelihood. This func-
tion uses incremental compilation to avoid
having to re-triangulate the entire junction
tree for each candidate network.
This score function and regular BIC scoring
(also augmented with a weighing parameter)
was tested for precision and inference time.
This analysis shows that there is a gain in
using the size of the junction tree as a com-
ponent in the scoring of learned nets, as these
net scored using this function was most often
faster when used for inference, and in some
cases even able to produce usable networks
where the networks learned with BIC-scoring
proved too complex.
Sprog | Engelsk |
---|---|
Udgivelsesdato | 2008 |
Udgivende institution | Aalborg universitet |