Reducing parameter value uncertainty in discrete Bayesian network learning: a semantic fuzzy Bayesian approach

Abstract

Bayesian network has gained increasing popularity among the data scientists and research communities, because of its inherent capability of capturing probabilistic information and reasoning with uncertain knowledge. However, the discrete Bayesian learning, with continuous and categorical variables, often shows poor performance because of parameter value uncertainty, arising due to strict boundary value of the discretized data and lack of knowledge on domain semantics. In this work, we propose semFBnet, a variant of Bayesian network with incorporated fuzziness and semantic knowledge, to reduce the uncertainty during parameter learning. The performance of semFBnet has been validated with prediction of daily meteorological conditions in two states of India, namely West Bengal and Delhi, for the years 2015 and 2016, respectively. The study of Dawid-Sebastiani score and the confidence interval analysis, in comparison with the state-of-theart and benchmark prediction techniques, demonstrate the effectiveness of the proposed semFBnet in reducing parameter value uncertainty.