http://lattes.cnpq.br/9956613423609786; BEZERRA, João Batista Nunes.
Abstract:
One of the challenges for the use of Bayesian networks is the construction of the Node
Probability Tables (NPT). The complexity for defining NPT is exponential, for large-scale
Bayesian networks, it is not feasible to manually define NPT. There are several techniques
that address this problem, among them, semiautomatic methods for the construction of NPT. The objective of this work is to measure and compare the modelling capability and accuracy of some of these methods: Weighted Sum Algorithm (WSA), Ranked Nodes Method (RNM), an adaptation of the Analytic Hierarchy Process (AHP) to the context of Bayesian networks, and a combination of WSA and AHP. Therefore, a case study with two units of analysis and a controlled experiment was performed. The accuracy of the methods was evaluated using historical data and simulated scenarios. Regarding the controlled experiment, a Randomized Complete Block Design (RCBD) was adopted. Ten software developers participated in the experiment and applied the methods to quantify uncertainties in a Bayesian network. These same developers manually defined NPT that served as a reference to measure the modelling capabilities and accuracy of the methods. The following measures of similarity were used to measure the modelling capacity of the methods: Brier Score (BS), Euclidean Distance (ED) and Mean Absolute Error (MAE). In order to measure the accuracy, the Accuracy Ratio (AR) of the methods was used. The main limitation of this study is its reduced scope of four methods due to the high costs involved in their application. According to the results obtained, it is possible to conclude that there are statistically significant differences between the methods. The method that presented the best result was the RNM, followed by the WSA, WSA-AHP and AHP.