Incremental Bayesian Network Learning for Scalable Feature Selection

摘要

Our aim is to solve the feature subset selection problem with thousands of variables using an incremental procedure. The procedure combines incrementally the outputs of non-scalable search-and-score Bayesian network structure learning methods that are run on much smaller sets of variables. We assess the scalability, the performance and the stability of the procedure through several experiments on synthetic and real databases scaling up to 139 351 variables. Our method is shown to be efficient in terms of both running time and accuracy.