Applying feature subset selection (FSS) methods on input data greatly reduces the dimensionality of a problem, thus making it manageable by machine learning methods. It helps already functioning learning systems to be faster and more accurate. It can also be used to discover the relationship between a large number of variables, and to use this knowledge to construct better models.
Approaching FSS by finding a Markov blanket guarantees that the excluded variables are unnecessary given the selected ones. Moreover, if we could find a Markov boundary, that would make the set of selected features minimal.
This Thesis investigates algorithms that can learn a Markov blanket or a Markov boundary from data. We examine them in detail, acknowledge their limits, and propose improvements to them.
We test these algorithms and their modified versions on a set of real data, and a set of artificial data in order to compare them. Based on this we evaluate them and make suggestions for further improvements.