D-separation is a term used to indicate that one set of nodes is conditionally independent of another set, given the current evidence.


A set of nodes A are d-separated from another set of nodes B, given evidence e if:

P(A|B,e) = P(A|e)


If sections of a network are d-separated from the queries we are calculating, we can ignore them. The algorithms used to calculate queries (inference) often detect D-separated nodes in order to ignore them.


The set of D-separated nodes changes depending on the evidence.


We can also ignore barren nodes, which when marginalized return a distribution of 1s which have no effect.