Remarks on learning hybrid Bayesian networks from data for medical research
F. D. Bookholt · Research Repository (Delft University of Technology) · 2013
In this thesis, we learn the structure of a hybrid Bayesian network (mixed continuous and discrete) for a small data set and learn the parameters. We segment the learning process in distinct steps and evaluate the possible options for each step. After a short introduction into the theory of Bayesian networks, we cover the topics: data preparation, algorithms, discretizing, tests for conditional independence, subnetworks, node ordering, directing arcs and learning parameters in sparse tables. After agreeing on a final quantified network, we compare our results to other networks modeling the same data and draw some conclusions about risk factors for and prediction op diabetes. The main conclusion is that learning hybrid Bayesian networks from a small data set is hard and requires much input from the modeler. Problems arise especially when testing for conditional independence in sparse frequency tables. Possible solutions are using different conditional independence tests, dividing the network into subnetworks, use a coarser discretization of the data and possible other options that we did not evaluate. The best option depends strongly on the data set you are working with, and a tailored solution must be made for each case.