An Object-Oriented Approach to Local Computation in Bayes Linear Belief Networks
Darren J. Wilkinson · COMPSTAT · 1998
A generalised conditional independence property based on zero partial correlations can be used to define Bayes linear graphical models, which may then be used as the basis for local computation. As for the full Bayesian case, local computation actually takes place on the clique-tree of the triangulated moral graph. An object-oriented framework for local computation is described which can be implemented via pure message-passing between objects representing adjacent clique-tree nodes. Messages consist of small matrices and vectors containing information about the current observation, making updating and propagation very fast. A sequential implementation, BAYES-LIN, is described.