Application of Bayesian Networks to Risk Assessment
Jidapa Kraisangka · D-Scholarship@Pitt (University of Pittsburgh) · 2019
Various approaches are used to estimate and predict risks. One of the most prevalent methods for risk assessment is the Cox's proportional hazard (CPH) model (Cox, 1972), a popular statistical technique used in risk estimation and survival analysis. The weaknesses of this approach are: (1) the underlying model can be only learned from data and is not readily amenable to refinement based on expert knowledge (2) the CPH model rests on several assumptions simplifying the interactions between the risk factors and the predicted outcome. While these assumptions are reasonable and the CPH model has been successfully used for decades, it is interesting to question them with a possible benefit in terms of model accuracy. This dissertation focuses on theoretical and practical aspects of risk assessment based on Bayesian networks (Pearl, 1988) as an alternative approach to the CPH model. The dissertation makes three contributions: (1) I propose a Bayesian network interpretation of the CPH (BN-Cox) model, a process of using existing CPH models as data sources for parameter estimation in Bayesian networks when original data are not available, and discuss methods for modeling such model computationally tractable (2) I empirically demonstrate in both context-sensitivity of the strength of influences of individual risk factors on the outcome variables in both Bayesian network model and the CPH model, and finally, (3) I propose and evaluate methods for enhancing the quality of Bayesian network parameters learned from small data sets, by means of priors.