Learning low inference complexity Bayesian networks

One of the main research topics in machine learning nowa- days is the improvement of the inference and learning processes in proba- bilistic graphical models. Traditionally, inference and learning have been treated separately, but given that the structure of the model conditions the inference complexity, most learning methods will sometimes produce ineficient inference models. In this paper we propose a new representa- tion for discrete probability distributions that allows eficiently evaluat- ing the inference complexity of the models during the learning process. We use this representation to create procedures for learning low infer- ence complexity Bayesian networks. Experimental results show that the proposed methods obtain tractable models that improve the accuracy of the predictions provided by approximate inference in models obtained with a well-known Bayesian network learner