摘要
潜变量模型在刻画因子间的相互关系以及因子与观测变量间的关联性方面具有重要作用.在实际应用中,观测数据往往呈现出重尾和极端值等特性.将经典的潜变量模型延伸到齐次隐马尔可夫模型,并建立了基于多元t-分布的极大似然统计分析程序.经验结果展示所建立的统计程序对消除异常点的影响是有效的.
Latent variable models play an important role in characterizing interrela- tionships among factor variables and constructing relationships between factors and observed variables. However, in real applications, data often have the heavy tails and/or contain extreme values. In this paper, we extend the classic latent variable model to the dynamic latent variable model mixed with homogenous hidden Markov model and establish maximum likelihood analysis procedure based on the multivariate t- distribution. The empirical results show that our proposed methodology is effective to down weight the influence of the outliers.
作者
夏业茂
陈高燕
刘应安
XIA Yemao CHEN Gaoyan LIU Yingan(School of Science, Nanjing Forestry University, Nanjing 210037)
出处
《系统科学与数学》
CSCD
北大核心
2016年第10期1783-1803,共21页
Journal of Systems Science and Mathematical Sciences
基金
国家自然科学基金(11471161)
南京市留学回国人员科技择优(013101001)资助课题
关键词
隐马尔可夫模型
潜变量模型
期望最大化算法
向前向后递推
Hidden Markov model； latent variable model； expectation-maximization；forward-backward recursion
