A Deterministic Annealing Approach to Learning Bayesian Networks

Graphical models have been very promising tools that can effectively model uncertainty, causal relation-ships, and conditional distributions among random variables. This paper proposes a new method for the induction of Bayesian Network structures from the data. The proposed method uses the concept of deterministic annealing; a global optimization approach originally developed for clustering and classification problems. In the proposed method the existence of an edge in the network is no longer considered as a hard 0/1 issue, but rather we assign a certain probability for the existence of an edge. The deterministic annealing procedure then proceeds by optimizing with respect to these edge probabilities. The experimental results show that the proposed approach achieves very promising results compared to other structure learning approaches.