Exploiting Data and Knowledge to Refine Bayesian Network Learning towards Tsunami Risk Analysis

Bayesian networks face the poor data quality issues, such as data infrequency,
data incompleteness, and non-uniform data types challenge, in automating a
tsunami risk analysis model. Since the analysis is highly domain-dependent, a
small amount of domain knowledge would significantly improve the model learning. In particular, the domain knowledge becomes more accessible with the
new development of large language models. In this paper, we propose a hybrid
knowledge-based and data-driven framework (HKDF) to learn Bayesian networks
for tsunami risk analysis. HKDF adapts Markov blanket (MB) techniques to
conduct data imputation while allowing domain knowledge to be embedded effectively in the entire BN learning process. The hybrid data-knowledge analytical solution becomes quite important when ubiquitous knowledge processing is commonly used in our daily lives. We evaluate the performance of HKDF over several benchmark data, state-of-the-art imputation and learning algorithms, and wellknown large language models. The experimental results show the effectiveness of our new framework.