基于贝叶斯网络的频繁模式兴趣度计算及剪枝

摘要

采用贝叶斯网络表示领域知识,提出一种基于领域知识的频繁项集和频繁属性集的兴趣度计算和剪枝方法BN-EJTR,其目的在于发现与当前领域知识不一致的知识,以解决频繁模式挖掘所面临的有趣性和冗余问题.针对兴趣度计算过程中批量推理的需求,BN-EJTR提供了一种基于扩展邻接树消元的贝叶斯网络推理算法,用于计算大量项集在贝叶斯网络中的支持度；同时,BN-EJTR提供了一种基于兴趣度阈值和拓扑有趣性的剪枝算法.实验结果表明,与同类方法相比,方法BN-EJTR具有良好的时间性能,而且剪枝效果明显；分析发现,经过剪枝后的频繁属性集和频繁项集相对于领域知识符合有趣性要求.%Based on background knowledge represented as a Bayesian network, this paper presents a BN-EJTR method that computes the interestingness of frequent items and frequent attributes, and prunes. BN-EJTR seeks to find inconsistent knowledge relative to background knowledge and to resolve the problems of un-interestingness and redundancy faced by frequent pattern mining. To deal with the demand of batch reasoning in Bayesian networks during computing interestingness, BN-EJTR provides a reasoning algorithm based on extended junction tree elimination for computing the support of a large number of items in a Bayesian network. In addition, BN-EJTR is equipped with a pruning mechanism based on a threshold for topological interestingness. Experimental results demonstrate that BN-EJTR has a good time performance compared with the same classified methods, and BN-EJTR also has effective pruning results. The analysis indicates that both the pruned frequent attributes and the pruned frequent items are un-interesting in respect to background knowledge.