Experiences with Modelling Issues in Building Probabilistic Networks

Abstract

Abstract. Building a probabilistic network for a real-life application is a difficult and time-consuming task. Methodologies for building such a network, however, are still lacking. Also, literature on network-specific modelling issues is quite scarce. As we have developed a large probabilistic network for a complex medical domain, we have encountered and resolved numerous non-trivial modelling issues. Since many of these issues pertain not only to our application but are likely to emerge for other applications as well, we feel that sharing them will contribute to engineering probabilistic networks in general.

Refining a Heuristic for Constructing Bayesian Networks from Structured Arguments

“…Since immoralities can result from following this guiding principle, the independence relations in the constructed BN graph should be verified manually, as prescribed by step 4 of the BR heuristic. This type of knowledge elicitation and verification is known to be a time-consuming and error-prone process in general [7]. Especially for larger or more densely connected BN graphs, it quickly becomes infeasible to verify all independence relations manually, as all possible chains for all possible combinations of instantiated variables need to be investigated.…”

“…Although the heuristic further suggests that the commonly used notion of causality be taken as a guiding principle [3], the resulting graph still has to be verified and refined in terms of the independence relations it represents. This type of knowledge elicitation is known to be timeconsuming [7], however, and moreover needs to be repeated for every adjustment to the original arguments. As a consequence, letting arc directions be set by a BN engineer is practically infeasible in investigative contexts such as police investigations, where evidence changes dynamically.…”

Matilda: A visual tool for modeling with Bayesian networks

“…However, although this may be the case for BN experts when the network at hand is small, it is far from clear how to proceed to someone who is not a BN expert. 6,11 Third, there are oracles that implement d-separation, for example, Wimberly's applet. 12 The problem is that simply telling the user that two nodes are d-separated is, again, not very helpful unless the user is a BN expert.…”

“…Here, the BN expert chooses the relationships to be investigated. This method is very time consuming, 11 as every situation needs a story tailored by the BN expert, with any change needing a new tailored story.…”

Constructing Bayesian Network Graphs from Labeled Arguments

“…BNs are well-suited for reasoning about the uncertain consequences that can be inferred from evidence. However, especially in data-poor domains, their construction needs to be done mostly manually, which is a difficult, time-consuming and error-prone process [7], and domain experts typically resort to using other tools such as argument diagrams, mind maps and ontologies [4,8]. Hence, we believe BN construction can be facilitated by automatically extracting information relevant for a BN from such tools.…”