Learning Bayes Nets using Score Functions and Dependency Constraints
By Oliver Schulte, Simon Fraser University
There are two well-established frameworks for learning Bayes nets: the "search
and score" paradigm and constraint-based learning. Score-based learning searches
for a graph structure G that maximizes a model selection score S(G,d) for a
given sample d. Constraint-based learning searches for a graph structure G that
entails the statistically significant dependencies and independencies
(correlations) in the sample d.
We propose a hybrid criterion for learning Bayes net structures that combines
search based on a scoring function S with information from statistical tests:
Search for a structure G that maximizes the score S, given the constraint that
the structure must entail the observed dependencies. We rely on the statistical
test only to accept conditional {\em dependencies}, not conditional
independencies. We show how to adapt local Bayes net search algorithms to
accommodate the observed dependencies. Simulation studies with GES and the BDeu
scoring function provide evidence that the additional dependency information
leads to an improved structure on small to medium sample sizes (e.g. $< 10^4$
data points with 10 variables).
Joint work with Wei Luo (Simon Fraser University) and Russ Greiner (University
of Alberta)