In this talk I look at models of conditional independence and their relationship to causal models. Studený's "imset" approach to representing conditional independence is reasonably well-known, but a 'dual' representation - supermodular functions - is less so. Supermodular functions have some appealing properties: testing for any particular conditional independence is easy, as is constructing new models by marginalising or conditioning. There is also a useful connection between supermodular functions and matroids.
