14.3 Constructing Bayesian Networks 343
understood. BN interfaces are discussed in subsection 14.3.3. In most cases,
BN development projects have no explicitly stated purpose. When there is
a stated purpose, it is usually too generic to be useful in the development
process. Having a detailed stated purpose would not entirely determine the
required accuracy, performance, and interfaces, but it would certainly help.
Because the amount of knowledge to be acquired may be very large, it
is important for the knowledge to be well organized. It is also important
to track the source of knowledge so that one can determine its trustworthi-
ness. These issues can be addressed to some degree by using ontologies. An
example of this is discussed in subsection 14.3.6.
Summary
The purpose of a BN should address these issues:
- Why the BN is being developed and how it will be used
- What will be covered by the BN
- Who will be using the BN
- The required accuracy, performance, and interfaces
14.3.2 Machine Learning
This subsection gives some background on current statistical methods for
constructing PDs. It begins with an overview of techniques for empirically
determining PDs, CPDs, and BNs from data. Such data are regarded as be-
ing used to “train” the probabilistic model, so the techniques are known as
machine learningmethods.
Machine learning is a very large area that would be difficult to survey ade-
quately, so we give only an overview. Since a BN is just a way of representing
a JPD, virtually any data-mining technique qualifies as a mechanism for con-
structing a BN. It is just a matter of expressing the end result as a BN. For
example, one might be interested in the body mass index (BMI) of individ-
uals in a research study. The individuals have various characteristics, such
as sex and age. Computing the average BMI of individuals with respect to
these two characteristics gives rise to a three-node BN as in figure 14.4. The
CPD for the BMI node gives the mean and standard deviation of the BMI for
each possible combination of sex and age in the study.
