Learning Bayesian network structure for risky behavior modelling
Bayesian Belief Networks (BBN) provide a comprehensible framework for representing complex systems that allows including expert knowledge and statistical data simultaneously. We explored BBN models for estimating risky behavior rate and compared several network structures, both expert-based and data-based. To learn and evaluate models we used generated behavior data with 9393 observations. We applied both score-based and constraint-based structure learning algorithms. The score-based structures represented better quality scores according to BIC and log-likelihood, prediction quality was almost the same for databased models and lower but sufficient for expert-based models. Hence, in case of limited data we can reduce computations and apply expert-based structure for solving practical issues.