Modeling multivariate continuous distributions is a task of central interest in statistics and machine learning with many applications in science and engineering. However, high-dimensional distributions are difficult to handle and can lead to intractable computations.
The Copula Bayesian Networks (CBNs) take advantage of both Bayesian networks (BNs) and copula theory to compactly represent such multivariate distributions. Bayesian networks rely on conditional independences in order to reduce the complexity of the problem, while copula functions allow modeling the dependence relation between random variables.
The goal of this thesis is to give a common framework for both domains and to propose new learning algorithms for copula Bayesian networks. To do so, we use the fact that CBNs have the same graphical language as BNs which allow us to adapt their learning methods to this model.
Moreover, using the empirical Bernstein copula both to design conditional independence tests and to estimate copulas from data, we avoid making parametric assumptions, which gives greater generality to our methods.