Priors on network structures. Biasing the search for Bayesian networks

Abstract

In this paper we show how a user can influence recovery of Bayesian networks from a database by specifying prior knowledge. The main novelty of our approach is that the user only has to provide partial prior knowledge, which is then completed to a full prior over all possible network structures. This partial prior knowledge is expressed among variables in an intuitive pairwise way, which embodies the uncertainty of the user about his/her own prior knowledge. Thus, the uncertainty of the model is updated in the normal Bayesian way.