Minimum redundancy maximum relevancy versus score-based methods for learning Markov boundaries

Feature subset selection is increasingly becoming an important preprocessing step within the field of automatic classification. This is due to the fact that the domain problems currently considered contain a high number of variables, and some kind of dimensionality reduction becomes necessary, in order to make the classification task approachable. In this paper we make an experimental comparison between a state-of-the-art method for feature selection, namely minimum Redundancy Maximum Relevance, and a recently proposed method for learning Markov boundaries based on searching for Bayesian network structures in constrained spaces using standard scoring functions.