Fast Shrinking parents-children learning for Markov blanket-based feature selection

High-dimensional data leads to degraded performance of machine learning algorithms and weak generalization of models, so feature selection is of great importance. In a Bayesian network (BN), the Markov blanket (MB) of a target node (T) is the best feature subset of that node. Therefore, this paper proposes Fast Shrinking parents-children learning for Markov blanket-based feature selection (FSMB), which first determines the parents-children of the target node, and then discovers spouses while checking candidate parents-children set. In spouse determination process, a secondary screening strategy is proposed to remove false-positive spouses effectively. In this process, once the spouses of T with respect to T’s child are determined, the parents-children set is immediately tested and the false-positive parents-children are removed in time, which not only can avoid the influence of false-positive parents-children on the subsequent spouse discovery, but also do not need to determine the spouses for false-positive parents-children. To verify the effectiveness of FSMB, experiments were performed on eight state-of-the-art MB algorithms on six standard networks and eight real datasets, the results show that FSMB outperforms other algorithms in terms of accuracy and efficiency.