项目名称: 条件独立结构的分解与学习
项目编号: No.11301408
项目类型: 青年科学基金项目
立项/批准年度: 2014
项目学科: 数理科学和化学
项目作者: 李本崇
作者单位: 西安电子科技大学
项目金额: 22万元
中文摘要: 条件独立是源自概率论的一个重要概念,它已被广泛地应用在现代统计学和人工智能的诸多领域。针对现代统计学中的高维数据,一个有效的策略是分解,一个有力的工具是计算代数几何。分解性是指把一个全局的统计问题转化成一系列局部问题并通过整合这些局部问题的结果来解决原来的全局问题。这一策略的理论基础是条件独立结构的分解。应用计算代数几何处理统计问题的优点在于,它可以解决以前未能有效解决的一些问题和现在出现的新问题。本项目中,我们首先研究一般条件独立结构的分解,主要是给出一般条件独立结构分解的一个定义,在此基础上研究其性质,并建立起条件独立结构分解与统计模型的可压缩性之间的关系。其次,条件独立结构的学习方面,我们主要考虑给定结构信息时,基于计算代数几何学的统计推断中的几个热门问题,包括图模型对应的消逝理想的基的计算,隐类模型参数的可识别性和参数真值是奇异点时似然比检验统计量的极限分布等问题。
中文关键词: 精确检验;忠实性;不平衡数据;VC 维数;高斯图模型
英文摘要: Conditional independence is an important concept originating from probabiity theory. It has been widely used in modern statistics and in many fields in artificial intelligence. To tackle high dimensional data in modern statistics, the property of decomposition is an effective strategy, and computational algebraic geometry is a strong tool. The property of decomposition means that one can split a global statistical problem into a series of local problems, and combine the results of these local problems to solve the original global problem. In fact, decomposition of conditional independence structures is the theoretical basis of this strategy. The advantages of using computational algebraic geometry to tackle statistical prolems are that it can solve some problems which can not be tackled effectively previously and some new problems. In this proposal, first, we study the decomposition of general conditional independence structures, that is, we mainly present a definition of decomposition for a general conditional independence structure, then, based on this definiton, we establish the relationship between decomposition and collapisibility of statistical models. Second, in the field of learning, based on structure information and tools in computational algebraic geometry, we focus on several popular problems of stat
英文关键词: Exact test;Faithfulness;imbalanced data;VC dimension;Gaussian graphical model