• 主题
高级搜索  >
历史搜索 清空历史
首页> 外文期刊>Statistics and computing >Introduction to 'Efficient local updates for undirected graphical models' by F. Stingo, G. Marchetti
【24h】

Introduction to 'Efficient local updates for undirected graphical models' by F. Stingo, G. Marchetti

机译：F. Stingo，G. Marchetti的“针对无向图形模型的有效本地更新”简介

获取原文
获取原文并翻译 | 示例
       
页面导航
  • 摘要
  • 著录项
  • 相似文献
  • 相关主题

摘要

Interconnectivity in systems can be represented by undirected graphs, and so graphical models naturally arise in many hierarchical Bayesian models, such as the neighbourhood structure in a segmented Gaussian Markov random field, or patterns of shared gene families across genomes, or friendship networks in Facebook. The paper by Stingo and Marchetti is motivated by a problem of inferring the interaction network for proteins from expression data. Inference over graph structures presents significant difficulties, particularly for MCMC sampling, because normalising constants usually depend on graph structure in an intractable way. For example, naive calculation for the segmented GMRF requires evaluating a determinant of the state-dependent precision matrix (concentration matrix in the present paper), which is generally computationally prohibitive for anything other than artifi-rndaily simplified problems. For the sub-class of decomposable graphs, the computational cost is made feasible by graph-theoretic results that provide a direct calculation of Cholesky factors, or equivalently the triangular decomposition. Stingo and Marchetti nicely summarise the theory for decomposable graphs and also dynamic algorithms over that class, describing ways that auxiliary data structures can be updated efficiently when using local moves. Their main contribution is a proposal for extending this machinery to general graphs. They show that by assuming a linear regression form for the likelihood, an approximated analysis over the general graphical model becomes tractable. Estimates presented for a protein network are informative, indicating that this paper makes a contribution to a topical, though notoriously difficult, problem in Bayesian modelling.
机译：系统中的互连性可以由无向图表示，因此图形模型自然会出现在许多分层的贝叶斯模型中，例如分段的高斯马尔可夫随机场中的邻域结构，或跨基因组的共享基因家族模式，或Facebook中的友谊网络。 Stingo和Marchetti的论文的动机在于从表达数据推断蛋白质相互作用网络的问题。对图结构的推断存在重大困难，尤其是对于MCMC采样而言，因为归一化常数通常以难处理的方式依赖于图结构。例如，对于分段GMRF的幼稚计算需要评估状态相关精度矩阵（本文中的浓度矩阵）的行列式，该矩阵通常在计算上对除人工简化问题以外的其他事情都是禁止的。对于可分解图的子类，通过图论结果可以使计算成本变得可行，图论结果可直接计算Cholesky因子，或等效地进行三角分解。 Stingo和Marchetti很好地总结了可分解图的理论以及该类的动态算法，描述了在使用局部移动时可以有效地更新辅助数据结构的方式。他们的主要贡献是将这种机制扩展到一般图形的建议。他们表明，通过对可能性采取线性回归形式，对一般图形模型的近似分析变得易于处理。蛋白质网络的估计值提供了有益的信息，表明本文为贝叶斯建模中的一个局部问题（尽管非常困难）做出了贡献。

著录项

  • 来源
    《Statistics and computing》 |2015年第1期|157-157|共1页
  • 作者

    Colin Fox;

  • 作者单位

    Department of Physics, University of Otago, Dunedin 9016, New Zealand;

  • 收录信息
  • 原文格式 PDF
  • 正文语种 eng
  • 中图分类
  • 关键词

相似文献

  • 外文文献
  • 中文文献
  • 专利
获取原文

客服邮箱：kefu@zhangqiaokeyan.com

京公网安备：11010802029741号 ICP备案号：京ICP备15016152号-6 六维联合信息科技 (北京) 有限公司©版权所有

  • 客服微信

  • 服务号