584 – Copulas: Past, Present, and Future
Efficient Learning of High Dimensional Copula Bayesian Networks
Yaniv Tenzer
Hebrew University
Gal Elidan
Hebrew University
Despite overlapping goals of multivariate modelling and dependence identification, until recently the fields of machine learning in general and probabilistic graphical models in particular have been ignorant of the framework of copulas. At the same time, complementing strengths of the two fields suggest the great fruitfulness of a synergy. In this talk we will give a taste of the of benefits that can arise from a symbiosis of the two fields. In particular, we will introduce a high-dimensional copula-based graphical model and then present a lightning-speed approach for automatically learning its structure from data. The approach is based on novel theory relating Spearman's Rho to the copula entropy, or expected likelihood of the model. No background in graphical models will be assumed.