Gaussian Bayesian network structure learning strategies based on canonical correlation analysis
Shuzhi Li, Guanghua Xu, Yongbao Feng · 2012
In order to solve the problem of low efficiency and low reliability of Gaussian Bayesian network structure learning methods, this paper proposes a new Gaussian Bayesian network structure learning algorithm from data based on the canonical correlation analysis. Firstly, by canonical correlation analysis of the son node and the candidate parent nodes, the correlation coefficients and correlation variables are given. Because the correlation coefficient indicates the association strength of family structure, we use correlation coefficient as measures of the family structure. Secondly, a new algorithm to establish parent nodes based on correlation variables is introduced. According to the correlation vectors to calculate the contribution value of candidate parent nodes, the contribution value is used to evaluate the association strength of parent node to son node. These nodes with the bigger contribution value are considered as the father nodes. Finally, the Bayesian network structure learning strategies is given based on canonical correlation analysis. The experimental results on the simulation standard data sets show that the new algorithm is effective and reliable.