Structure Learning of Bayesian Networks Based On the LARS-MMPC Ordering Search Method
Chuchao He, Xiaoguang Gao · 2018
A given ordering among variables can significantly improve the accuracy of learning in Bayesian network structures. In this study, we propose using a combined Least Angle Regression (LARS) and Max-Min Parent and Children (MMPC) algorithm based on known root nodes specified by domain experts in order to obtain the optimal ordering. First, with a fixed root node, a partial ordering is tailored from the entire ordering by using the LARS algorithm. A further sequence is then obtained by combining all the different partial orderings. Parent and children sets are detected among the remaining nodes by the MMPC algorithm. Finally, a complete ordering is derived from the sequence and the parent and children sets, and the optimal structure is learnt by the K2 algorithm based on the ordering. Experiments showed that compared with other competitive methods, the proposed algorithm performed well in terms of balancing the learning accuracy with time consumption.