**Description** |
Bayesian Networks (BNs) are popular models for representing probabilistic knowledge,
and reasoning with uncertainty. Automatic construction of BNs from a set of
example data is an active area of research. In many applications of BNs, we are
mainly interested in predicting the value of one particular variable (the class) from
the remaining variables (the attributes). Such BNs are called Bayesian Network
Classifiers (BNCs). To fit a single BNC to a data set, potentially time consuming
iterative optimization
algorithms are required. To find out which of the many possible BNCs works the best,
many different BNCs have to be fitted to the data. Therefore it is important that
fitting a single BNC is performed as efficiently as possible.
In this project, a new algorithm for learning BNCs
from data has been developed and implemented, and its efficiency has been compared
experimentally to existing algorithms for performing this task.
This comparison was performed by applying the algorithms to benchmark data sets from
the UCI machine learning repository, and to artificially generated data.
The results showed that the newly developed algorithm is, in many cases, more
efficient than its competitors,
in particular when the BNC to be fitted has a relatively simple structure. |