Using Survey and Weighted Functions to Generate Node Probability Tables for Bayesian Networks

Abstract

Recently, Bayesian networks became a popular technique to represent knowledge about uncertain domains and have been successfully used for applications in various areas. Even though there are several cases of success and Bayesian networks have been proved to be capable of representing uncertainty in many different domains, there are still two significant barriers to build large-scale Bayesian networks: building the Directed Acyclic Graph (DAG) and the Node Probability Tables (NPTs). In this paper, we focus on the second barrier and present a method that generates NPTs through weighted expressions generated using data collected from domain experts through a survey. Our method is limited to Bayesian networks composed only of ranked nodes. It consists of five steps: (i) define network's DAG, (ii) run the survey, (iii) order the NPTs' relationships given their relative magnitudes, (iv) generate weighted functions and (v) generate NPTs. The advantage of our method, comparing with existing ones that use weighted expressions to generate NPTs, is the ability to quickly collect data from domain experts located around the world. We describe one case in which the method was used for validation purposes and showed that this method requires less time from each domain expert than other existing methods.