Constructing dependence ordering for B&B technique in learning Bayesian belief network
Wei Gao, Kun Niu · 2011
The data mining technology is more and more widely used. How to construct a Bayesian belief network has been discussed in many different ways. As one of the classical algorithms, the branch and bound technique based on the minimum description length principle has been proposed by Joe Suzuki in 1998. But one of the most important premises of the B&B Technique is that an attributes' dependence ordering has been prepared. To address the problem, a new method is proposed for attributes' dependence ordering. The algorithm first constructs a dependence tree using training dataset, then we use breadth first searching and get the dependence ordering. That results in a raw ordering. Then we order the nodes that are in the same layer of the dependence tree in order to make the result more accurate. This paper uses real datasets from the telecom industry as the test datasets. The result shows that the algorithm can construct the dependence ordering with good performance.