The propagation-based approximation algorithm is a more generalized version of the belief propagation algorithm and works on the same principle of passing messages. In the case of exact inference, we used to construct a clique tree and then passed messages between the clusters. However, in the case of the propagation-based approximation algorithms, we will be performing message passing on cluster graphs.
Let's take the simple example of a network:
It is possible to construct multiple cluster graphs for this network. Let's take the example of the following two cluster graphs:
Fig 4.2 shows two possible cluster graphs for the network in Fig 4.1. The cluster graph in Fig 4.2(a) is a clique tree and the clusters are (A, B, C) and (B, C, D). Whereas, the cluster graph in Fig 4.2(b) has four clusters (A, B), (B, C), (C, D), and (A, D). It also has loops: