Connectable and Independent Junction Tree-Based Compilation Technique of Object-Oriented Bayesian Networks

Object-oriented Bayesian network (OOBN) is a method for building compositional and hierarchical Bayesian network (BN) models that promote reuse and simple maintenance. Reasoning with both BNs and OOBNs entails the computational job of inference, the computation of new posterior probability distributions based on a set of evidence. A widely used inference strategy in conventional BN is to compile the BN into a junction tree (JT) before conducting standard inference. In the case of OOBN, it is first flattened into the underlying BN before performing the JT-based compilation. However, large OOBNs flatten to complex and larger BNs can be computationally intensive to compile into JTs due to the complexity of compilation being exponential to the size of BNs. To cope with these performance issues, techniques like Incremental Compilation (IC) avoid reconstructing JT from scratch after each modification of a BN. However, none of the existing works were able to reduce the computational complexity of compilation. Hence, in this paper, we propose a new compilation algorithm that compiles the OOBN without flattening it and re-using the existing JTs of embedded components of the OOBN. Evaluation results show that our proposed algorithm effectively reduces the computation time for JT construction of OOBN.