Discovery of context-specific Markov blankets

The notion of context-specific Markov blankets (CSMB) is a refinement of Markov blankets (MB). An important property of context specific Markov-blankets is that in the worst case they need as many parameters as a model using a standard Markov blanket, but frequently considerably fewer. The result is expected to be a much improved probabilistic model of the data in practice. An algorithm for discovery of the CSMB is presented, and empirical results show that it is capable of successfully recovering context specific (''local") structures.