一种具有缺失数据的贝叶斯网络结构学习方法

在数据缺失的情况下,文章建立的贝叶斯网络结构学习方法BC-ISOR基于界定折叠方法进行变量集联合概率的估计,基于依赖分析的基本思想进行贝叶斯网络的结构学习。该方法的概率估计与数据的缺失率无关,并有效降低了条件独立性检验的次数和阶数,因而具有良好的学习效率并能避免陷入局部最优;针对Asia网络的实验结果表明,该方法比经典方法SEM具有更好地时效性和精确性。

A method for learning Bayesian networks from missing data

A new method BC-ISOR is put forward for learning Bayesian networks from missing data in this paper.The method uses the underlying principle of the bound and collapse method to estimate the joint probability of variable sets,and it adopts basic ideas of dependency analysis to learn Bayesian networks.The probability estimation of the method is independent of the percentage of missing data.The method greatly reduces the number and order of conditional independence tests.So the method has good efficiency and avoids stopping at the locally optimal network.The experimental results on the Asia network show that the method BC-ISOR is more efficient and more accurate than the classical method SEM.