L1 regularized ordering for learning Bayesian network classifiers

Ying Wang, Hao Wang, Kui Yu, Hongliang Yao

Abstract: Learning a Bayesian network classifier from data is an active research topic in data mining. The key problem for constructing a Bayesian network classifier is to learn an accurate Bayesian network structure which is a difficult task. The K2 algorithm, as one of the most efficient Bayesian network learning methods can deal with this difficult task. However, K2 requires a variable ordering in advance. Existing methods for establishing this ordering neglect information of the variables selected. To address this problem, in this paper, we propose an L1 regularized Bayesian network classifier (L1-BNC). L1-BNC defines a variable ordering by the LARS (Least Angle Regression) method, and then with this ordering it uses K2 to construct a Bayesian network classifier. In comparison with seven Bayesian network classifiers, L1-BNC outperforms those classifiers on most datasets.

External IDs:dblp:conf/icnc/WangWYY11