贝叶斯网边际马尔科夫子图及其应用

Marginal Markov Subgraph of Bayesian Network and Its Applications

贝叶斯网络利用有向无圈图对多元联合概率分布中条件独立性进行约束,以实现其在不确定推理中的模块化分解,降低概率推理的计算复杂度.它在概率推理、机器学习和因果推理中都有广泛的应用.在实际中,如果采用分而治之或模型压缩的方法对贝叶斯网络进行结构学习或统计推断,那么需要人们寻找边际分布的极小马尔科夫子图(或极小独立图)来建立边际模型.为此,本文基于贝叶斯网的道义图研究贝叶斯网边际模型的极小马尔科夫子图,从统计和图论的观点对其进行了细致的刻画.针对DAG模型的可压缩性,本文将基于有向导出路径的性质给出更直观的等价条件,同时又给出了若干充分条件,这为判断模型是否可压缩到局部子模型上提供了更多的理论工具.

Bayesian networks utilize directed acyclic graphs(DAGs)to constrain conditional independencies in multivariate joint probability distribution,so as to realize its modular decomposition in uncertainty reasoning and reduce the computational complexity of probabilistic reasoning.They are widely used in probabilistic reasoning,machine learning and causal inference.In practice,if structure learning or statistical inference was performed by adopting the idea of dividing and conquering or model collapsing,we have to establish the marginal models by finding their minimal Markov subgraphs(or minimal independence maps).Therefore,this paper details minimal Markov subgraphs for marginal models of Bayesian networks,and provides the refined characterization on them from the perspectives of statistics and graph theory.For the collapsibility of DAG,this paper gives more intuitive equivalent conditions based on the properties of directed inducing paths,and also proposes some sufficient conditions,which provides more theoretical tools for judging whether the considered models can be collapsible onto local sub-models.