两两关系马尔科夫网的自适应组稀疏化学习

稀疏化学习能显著降低无向图模型的参数学习与结构学习的复杂性,有效地处理无向图模型的学习问题.两两关系马尔科夫网在多值变量情况下,每条边具有多个参数,本文对此给出边参数向量的组稀疏化学习,提出自适应组稀疏化,根据参数向量的模大小自适应调整惩罚程度.本文不仅对比了不同边势情况下的稀疏化学习性能,为了加速模型在复杂网络中的训练过程,还对目标函数进行伪似然近似、平均场自由能近似和Bethe自由能近似.本文还给出自适应组稀疏化目标函数分别使用谱投影梯度算法和投影拟牛顿算法时的最优解,并对比了两种优化算法进行稀疏化学习的性能.实验表明自适应组稀疏化具有良好的性能....

Full description

Saved in:
Bibliographic Details
Published in自动化学报 Vol. 41; no. 8; pp. 1419 - 1437
Main Author 刘建伟 任正平 刘泽宇 黎海恩 罗雄麟
Format Journal Article
LanguageChinese
Published 中国石油大学自动化研究所 北京 102249%中国科学院软件研究所基础软件国家工程研究中心 北京 100190 2015
Subjects
两两马尔科夫网
无向图模型
稀疏化学习
自适应组稀疏化
无向图模型
Undirected graphical models
pairwise Markov network
稀疏化学习
adaptive group sparsity
sparse learning
两两马尔科夫网
自适应组稀疏化
Online AccessGet full text
ISSN0254-4156
1874-1029
DOI10.16383/j.aas.2015.c140682

Cover

More Information
Summary:稀疏化学习能显著降低无向图模型的参数学习与结构学习的复杂性,有效地处理无向图模型的学习问题.两两关系马尔科夫网在多值变量情况下,每条边具有多个参数,本文对此给出边参数向量的组稀疏化学习,提出自适应组稀疏化,根据参数向量的模大小自适应调整惩罚程度.本文不仅对比了不同边势情况下的稀疏化学习性能,为了加速模型在复杂网络中的训练过程,还对目标函数进行伪似然近似、平均场自由能近似和Bethe自由能近似.本文还给出自适应组稀疏化目标函数分别使用谱投影梯度算法和投影拟牛顿算法时的最优解,并对比了两种优化算法进行稀疏化学习的性能.实验表明自适应组稀疏化具有良好的性能.
Bibliography:Undirected graphical models, pairwise Markov network, sparse learning, adaptive group sparsity
Sparse learning can significantly reduce the complexity of parameter learning and structure learning and effectively deal with learning problems of undirected graphical models. In the case of pairwise Markov network, in which each variable has more than two values, the number of parameters associated with an edge is more than one. This paper proposes a group sparse learning approach for the parameters associated with edges, and puts forward an adaptive group sparse learning algorithm, which can adaptively adjust the degree of penalty according to the norm of the parameters vector. This paper compares the performance of sparse learning using different edge potentials. In order to speed up the training process, three approximate object functions are given, including pseudo likelihood approximation, mean field approximation and Bethe free energy approximation. Two optimization algorithms, i.e., projected quasi-Newton al
ISSN:0254-4156
1874-1029
DOI:10.16383/j.aas.2015.c140682