Junction tree decomposition for parallel exact inference

We present a junction tree decomposition based algorithm for parallel exact inference. This is a novel parallel exact inference method for evidence propagation in an arbitrary junction tree. If multiple cliques contain evidence, the performance of any state-of-the-art parallel inference algorithm achieving logarithmic time performance is adversely affected. In this paper, we propose a new approach to overcome this problem. We decompose a junction tree into a set of chains. Cliques in each chain are partially updated after the evidence propagation. These partially updated cliques are then merged in parallel to obtain fully updated cliques. We derive the formula for merging partially updated cliques and estimate the computation workload of each step. Experiments conducted using MPI on state-of-the-art clusters showed that the proposed algorithm exhibits linear scalability and superior performance compared with other parallel inference methods.