基于时序隐变量模型的因果关系发现算法

Causal discovery algorithm based on latent factor models for time series data

为在基于隐变量模型的因果关系发现算法中综合考虑隐变量之间的瞬时性和延时性因果效应,构建以动态贝叶斯网络为基础的时序隐变量模型,提出对应的因果关系发现算法。使用因子分析的方法估计测量模型中的因子载荷矩阵,应用结构向量自回归模型估计自回归矩阵,利用数据的非高斯性依次学习模型中隐变量之间的瞬时效应矩阵与延时效应矩阵,构建时序隐变量模型的因果网络结构。实验结果验证了算法的有效性。

To comprehensively consider the instantaneous and time-lag causal effects between hidden variables in the causal discovery algorithms based on the latent factor model,a latent factor model for time series data(LFTS)based on dynamic Baye-sian networks was presented with its corresponding algorithm.The factor analysis method was used to estimate the factor loading matrix in the measurement model,and the structure vector autoregressive model was utilized to estimate the autoregressive matrix.The non-Gaussianity of the data were employed to sequentially learn the instantaneous effect matrix and the time-lag effect matrix between latent variables,and the causal structure of the LFTS was obtained.Experimental results verify the effectiveness of the algorithm.