Abstraction in Markov Networks

In this paper a new approach is presented for taming the complexity of performing inferences on Markov networks. The approach consists in transforming the network into an abstract one, with a lower number of vertices. The abstract network is obtained through a parti- tioning of its set of cliques. The paper shows under what conditions exact inference may be obtained with reduced cost, and ways of partitioning the graph are discussed. An example, illustrating the method, is also described.