Context-specific independence in Bayesian networks

摘要: Bayesian networks provide a language for qualitatively representing the conditional independence properties of distribution, This allows natural and compact representation eases knowledge acquisition, supports effective inference algorithms. It is well-known, however, that there are certain independencies we cannot capture within network structure: hold only in contexts, i.e., given specific assignment values to variables, In this paper, propose formal notion context-specific (CSI), based on regularities probability tables (CPTs) at node. We present technique, analogous (and on) d-separation, determining when such holds network. then focus particular qualitative scheme--tree-structured CPTs-- capturing CSI. suggest ways which can be used support algorithms, particular, structural decomposition resulting improve performance clustering an alternative algorithm outset conditioning.