Compression of Bayesian Networks with NIN-AND Tree Modeling

We propose to compress Bayesian networks (BNs), reducing the space complexity to being fully linear, in order to widely deploy in low-resource platforms, such as smart mobile devices. We present a novel method that compresses each conditional probability table (CPT) of an arbitrary binary BN into a Non-impeding noisy-And Tree (NAT) model. It achieves the above goal in the binary case. Experiments demonstrate that the accuracy is reasonably high in the general case and higher in tested real BNs. We show advantages of the method over alternatives on expressiveness, generality, space reduction and online efficiency.