小数据集条件下的多态系统贝叶斯网络参数学习

针对贝叶斯网络中多父节点条件概率分布参数学习问题,提出了一种适用于多态节点、模型不精确、样本信息不充分情形的参数学习方法.该方法利用因果机制独立假设,分解条件概率分布,使条件概率表的规模表现为父节点个数和状态数的线性形式;利用Leaky Noisy-MAX模型量化了多态系统模型未含因素对参数学习的影响;从小样本数据集中获取模型参数并合成条件概率表.结果表明,该方法能提高参数学习效率与精度.

Parameters Learning of Bayesian Networks for Multistate System with Small Sample

To learn parameters for conditional probability distribution of multi-father nodes,a method was proposed which applys to the multistate nodes in the inaccurate model under the condition of insufficient sample information.Using the assumption of independence of causal interaction,the conditional probability distribution is decomposed and the size of conditional probability table is linear in the numbers of the parent nodes and their states.Using Leaky Noisy-MAX model,the influence of factors not included in the multistate system model can be quantified on the parameters learning.The model parameters extracted from small sample can create conditional probability tables.The results show that the method can improve the efficiency and precision of parameter learning.