Title: Composite Bayesian inference

Abstract: This paper revisits the concept of composite likelihood from the perspective of probabilistic inference, and proposes a generalization called "super composite likelihood" for more flexible use of data information. It is shown that super composite likelihood yields a class of discriminative models suitable for unsupervised or weakly supervised learning.