A Conservative Feature Selection Algorithm with Informatively Missing Data

Alexandre Aussem (1) , Sergio Rodrigues de Morais (1)

1 LIESP - Laboratoire d'Informatique pour l'Entreprise et les Systèmes de Production

This papers introduces a novel conservative feature subset selection method with informatively missing data, i.e., when data is not missing at random but due to an unknown censoring mechanism. This is achieved in the context of determining the Markov blanket (MB) of the target variable in a Bayesian network. The method is conservative in the sense that it constructs the MB that reflects the worst-case assumption about the missing data mechanism, when the missing values cannot be inferred from the available data only. An application of the method on synthetic and real-world incomplete data is carried out to illustrate its practical relevance.

Mots clés: Bayesian networks; missing data; Markov boundary; probabilistic classification; feature subset selection

Domaines: Apprentissage [cs.LG]

Alexandre Aussem, Sergio Rodrigues de Morais. A Conservative Feature Selection Algorithm with Informatively Missing Data. Journées Francophone sur les Réseaux Bayésiens, May 2008, Lyon, France. ⟨hal-00280404⟩