Probabilistic graphical Markov model learning: an adaptive strategy

In this paper an adaptive strategy to learn graphical Markov models is proposed to construct two algorithms. A statistical model complexity index (SMCI) is defined and used to classify models in complexity classes, sparse, medium and dense. The first step of both algorithms is to fit a tree using the Chow and Liu algorithm. The second step begins calculating SMCI and using it to evaluate an index (EMUBI) to predict the edges to add to the model. The first algorithm adds the predicted edges and stop, and the second, decides to add an edge when the fitting improves. The two algorithms are compared by an experimental design using models of different complexity classes. The samples to test the models are generated by a random sampler (MSRS). For the sparse class both algorithms obtain always the correct model. For the other two classes, efficiency of the algorithms is sensible to complexity