Markov Blanket Based Feature Selection: a Review of Past Decade
Shunkai Fu, Michel C. Desmarais · PolyPublie (École Polytechnique de Montréal) · 2010
Abstract—This paper summarizes the related works about feature selection via the induction of Markov blanket which can be traced back to 1996, and the concept of Markov blanket itself firstly appeared even earlier in 1988. Our review not only covers a series of published algorithms, including KS, GS, IAMB and its variants, MMPC/MB, HITON-PC/MB, Fast IAMB, PCMB and IPC-MB (ordered as their appearing time), but why they were invented and their relative advantage as well as disadvantages, from both theoretical and practical viewpoint. Besides, it is noticed that all of these mentioned works are all constraint learning which depends on conditional independence test to induce the target, instead of via score-andsearch, another mainstream manner as applied in the structure learning of one closely related concept, Bayesian network. Bing the first one, we discuss the cause which uncovers that this choice is not accidental, though not in a formal way. The discussion covered here is believed a valuable reference for academic researchers as well as applicants. Index Terms—Feature selection, Markov Blanket. I.