基于贝叶斯网络的马尔科夫毯预测学习

LEARNING MARKOV BLANKET PREDICTION BASED ON BAYESIAN NETWORK

研究变量之间的预测能力在许多领域都有重要意义,通过这种研究,能够揭示变量之间的制约机制,贝叶斯网络是研究变量之间预测能力的有力工具。本文使用依赖分析方法建立基于贝叶斯网络的马尔科夫毯预测,其核心问题是贝叶斯网络结构学习。目前,基于依赖分析的贝叶斯网络结构学习方法主要存在三个问题:(1)需要进行大量的高维条件概率计算,(2)容易丢失弱联合依赖边,(3)对边的方向的确定具有局限性。针对这些问题,本文提出了首先进行递推条件独立性检验,然后进行因果语义定向,最后进行冗余边检验的贝叶斯网络结构学习方法。该方法能够有效地避免这些问题,更准确地建立马尔科夫毯预测。

Research on prediction ability between variables is very important in many domains. Through this kind of research, the restriction mechanism between variables can be revealed. Bayesian network is a powerful tool of studying prediction ability between variables. In this paper, Markov blanket prediction based on Bayesian network is built by means of dependency analysis. The key of learing Markov blanket prediction is to set up Bayesian network structure. At present, there mainly exist three problems in the methods of learning Bayesian network structure based on dependency analysis: (1) it is needed to calculate a large number of conditional probability with high dimension, (2) some weak joint dependent relationships are easily lost, (3) there are often some edges that can not be oriented. To solve these problems, a new method of learning Bayesian network structure is presented. First conditional independence between variables is tested in the form of recursion, then the edges are oriented according to the causal semanitics of an are' s direction, at last redundancy edges are checked up. This kind of method can effectively avoid these problems to build Markov blanket prediction more accurately.