Data discretization and Bayesian network modeling: A case study in R

Bayesian networks (BNs) are a graphical representation of a probability distribution over a set of variables. It comprises a directed acyclic graph (DAG) and a set of probability distributions. The first aim of this thesis is to review and introduce the BNs and data discretization. The second one is to illustrate how discrete BNs are constructed from the data which include continuous variables thorough discretization. The third one is to compare the effect of the two commonly-used data discretization approaches on the BNs. The last one is to study the relationships among the Beck Depression Inventory, Beck Hopelessness Scale, Rosenberg Self-Esteem Scale scores and demographic and socio-economic variables with BN modeling. The data of 823 university students consist of 21 continuous and discrete relevant psychiatric, demographic and socio-economic variables The continuous variables are discretized by using the Information-Preserving Discretization (IPD) and domain knowledge available in the literature and, consequently, two discrete BNs are learned from the data sets and constructed in statistical software R and the results are presented via figures and probabilities. One of the most significant results is that the structure of the two BNs does not significantly differ. The only difference is that in the first Bayesian network model, the gender of the students influences the level of depression, with female students being more likely to be more depressive, while in the second model, social activity directly influences the level of depression and the presence of social activity decreases the risk of being more depressive. Another important result is that in each model, depression influences both the level of hopelessness and self-esteem in students; additionally, the level of depression has a positive impact on the level of hopelessness, but a negative impact on the level of self-esteem. The last remarkable result is that, based on the BIC values and specificities of the two models, the BN whose continuous data were discretized by the IPD outperforms the BN whose data were discretized by the domain knowledge.