The Discrete-Choice Copula Bayesian Network Model

Bayesian network classifiers can be combined with copulas to accurately model continuous random variables. However, applications of such frameworks often restrict the graph structure or the type of variables that may be included. We develop the predictive copula Bayesian network (PCBN), a probabilistic classification framework which allows one to model continuous, ordinal discrete and categorical variables explicitly, while allowing variables to be correlated via copulas. We also introduce an algorithm which constructs a PCBN based on (predictive) scores, without assuming tree-like structures. We apply the framework to both simulated and empirical travel mode data and show that PCBNs can accurately model the data, especially when the data is highly correlated.