贝叶斯网络结构稀疏学习研究进展

摘要

贝叶斯网络结构稀疏化学习因其既能简化结构又能保留原始网络中的重要信息，已经成为当前贝叶斯网络的研究热点。文中首先讨论贝叶斯网络结构稀疏学习的必要性、贝叶斯网络稀疏性的定义，并在此基础上介绍现有的贝叶斯网络结构稀疏学习研究思路。然后，回顾一般的贝叶斯网络结构学习方法，并分析它们在高维背景下存在的问题，进而发现基于评分的方法通常适合于贝叶斯网络结构的稀疏学习，因此重点介绍贝叶斯网络结构稀疏学习的目标函数和优化求解算法。最后，探讨未来贝叶斯网络结构稀疏学习的一些研究方向。%Sparse structure learning of Bayesian networks can simplify network structure without losing important information of the original network structure. In this paper, the necessity of the sparse structure learning of Bayesian networks and the definition of the sparsity of those are firstly discussed. Based on the general structure learning of Bayesian networks, the existing problems for high-dimensional data are analyzed, and then it is found that score-based structure learning is suitable for sparse structure learning. Therefore, the objective functions and their optimization algorithms are mainly described. Finally, some meaningful research trends are discussed.