De-Causalizing NAT-Modeled Bayesian Networks for Inference Efficiency

摘要

Conditional independence encoded in Bayesian networks (BNs) avoids combinatorial explosion on the number of variables. However, BNs are still subject to exponential growth of space and inference time on the number of causes per effect variable in each conditional probability table (CPT). A number of space-efficient local models exist that allow efficient encoding of dependency between an effect and its causes, and can also be exploited for improved inference efficiency. We focus on the Non-Impeding Noisy-AND Tree (NIN-AND Tree or NAT) models due to its multiple merits. In this work, we develop a novel framework, de-causalization of NAT-modeled BNs, by which causal independence in NAT models can be exploited for more efficient inference. We demonstrate its exactness and efficiency impact on inference based on lazy propagation (LP).