中文摘要：

贝叶斯网是不确定性问题知识表达和推理中最重要的一个理论模型．迄今为止人们提出了许多贝叶斯网结构学习算法，基于约束满足和评分搜索相结合的混合方法是其中的一个研究热点．以I—B＆B—MDL为基础，提出了一种快速的学习算法．新算法不仅利用约束知识来压缩搜索空间，而且还用它作为启发知识来引导搜索．首先利用0阶和少量的1阶测试有效地限制搜索空间，获得网络候选的连接图，减少了独立性测试及对数据库的扫描次数，然后利用互信息作为启发性知识来引导搜索，增加了B＆B搜索树的截断．在通用数据集上的实验表明：快速算法能够有效地处理大规模数据，且学习速度有较大改进．

英文摘要：

Bayesian network （BN） is one of the most important theoretical models for uncertainty knowledge expression and reasoning. So far, many BN structure learning algorithms have been proposed. In this paper, a fast algorithm FI-B＆B-MDL is developed, which considerably speeds up the original I- B＆B-MDL algorithm. Unlike I-B＆B-MDL, the new FI-B＆B-MDL first uses only order-0 and a small number of order-1 independence tests to obtain an original structure graph so that the number of independence tests and database passes can be decreased, and then takes mutual information between nodes as the heuristic knowledge to lead MDL searches so that the cut-offs of B ＆ B search trees can be increased, and consequently the search process is accelerated. Experimental results show that the new algorithm is effective and efficient in large scale databases, and it is faster than the original algorithm.