16.4 Ontologies for Bayesian Networks 375
The CPDs are the most complex elements. In general, a CPD is a list of PDs.
The list is contained in aDPISelement. PDs are specified byDPIelements.
If a node has no incoming edges, then its CPD is a PD and there is only a
singleDPIelement. Nodes with incoming edges must specify several PDs.
The published DTD for XBN does not support continuous random variables,
so it was necessary to add two attributes to theDPIelement: theMEANand
VARIANCE.
The XBN format has a number of limitations as the basis for the BW. In its
current published form, it only supports random variables with a finite num-
ber of values. It does not support continuous random variables. It should be
possible to specify a wide variety of types of PD. Another significant lim-
itation is its lack of a mechanism for referring to external resources or for
external documents to refer to the BN. This makes it difficult to use this
mechanism to satisfy the requirement for common variables, and there is
only limited support for annotation.
These considerations suggest that a better choice of language for the BW
is OWL. We now present a series of three OWL ontologies that satisfy the
requirements for the BW. We present them in top-down fashion, starting with
high-level concepts and successively elaborating them:
- The ontology of phenomena which can be modeled using BNs
- The ontology of networks of CPDs
- The ontology of elementary PDs
Figure 16.2 Ontology for Bayesian networks.