Variational message passing

If the model is in the conjugate exponential family, the VB-EM algorithm can be implemented as the variational message passing (VMP) algorithm ([19]). The algorithm is based on local computations, in which each factor is represented by a node. When a node is updated, it receives messages from its children and parents and uses those messages to compute the new approximate distribution and relevant expectations. The advantage of this message passing formulation is that the computations are local and depend on well-defined messages. This makes it easy to modify the model by adding or changing nodes.