Abstract: Inferring the causality among variables using Bayesian networks has been applied widely in the field of artificial intelligence. The algorithms for constraint-based of constructing Bayesian networks usually return the Markov equivalent class of the real network from observed data, which cannot infer causality effectively because of the existence of undirected edges. In order to improve the inference of Bayesian networks, a model merging strategy combining the Bayesian network score function and the ensemble learning is proposed to reduce the number of undirected edges by integrating multiple Bayesian networks. The experimental results show that it can reduce the number of undirected edges apparently by merging weighted network structures and improve the accuracy of the final network structure as well, which validates the effectiveness of the algorithm.

Key words: Bayesian networks, score function, model merging, causal inference

摘要： 利用贝叶斯网络进行因果关系推理已广泛应用于人工智能领域。基于约束方法从观测数据中构建贝叶斯网络通常得到的是其马尔科夫等价类，因存在无向边而无法进行有效的因果推断。为此，基于贝叶斯网络评分函数，并结合集成学习提出了一种模型融合算法，通过对不同的网络结构加权融合，以减少网络中无向边的个数，进而提高其可推断性。实验结果表明，不仅显著减少了无向边条数，也提高了最终网络结构的学习效果，验证了算法的有效性。

关键词: 贝叶斯网络, 评分函数, 模型融合, 因果推断