分组样本下Bayes网络条件概率的学习算法

Learning Algorithms of Bayesian-Network Conditional Probabilities from Grouped Sample

本文应用带盘的 Bayes网络作为分析模型 ,对于学习实例数据库为分组样本的关于各组样本私有条件概率的学习算法进行讨论 ,构建出两层学习结构 :第一层针对各组私有条件概率分布 Θij(l) s的学习 ;第二层针对是各组公有条件概率分布Θij的学习 .算法在在综合公有后验条件概率分布和本组学习实例数据的基础上 ,实现对各组私有条件概率分布的学习 ,并可以通过调整组间差异性信度

In this paper, we applied Bayesian network with plates as analysis model to propose a learning algorithm that can derive private conditional probabilities from grouped sample. We have exploited a two level learning architecture: at first level, we have a set of parameters that express private conditional probabilities, i.e. the Θ ij (l)s; at the second level, we have a set of parameters that express popular conditional probabilities, i.e. the vectorΘ ij . The overall result is that the posterior private conditional probabilities are a compromise between pecific information and the posterior distribution of the population, allowing each posterior private conditional probability to borrow value from all the experience contained in the database. The degree of compromise can be modulated by tuning the parameterβ.