| 摘要: |
| 为了使统计分析有效进行,特征筛选问题在超高维领域已被众多学者广泛研究;针对现存特征筛选方法不
能灵活处理超高维纵向数据的组内相关性问题,提出一个基于动态协方差建模的迭代特征筛选方法,并称之为迭代的动态特征筛选方法;在每次迭代过程中,均使用修正的 Cholesky 分解代替静态协方差矩阵建模方法对纵向数据的组内协方差矩阵进行动态建模,获得灵活的组内协方差矩阵估计,然后将所得估计代入广义估计方程中,并基于广义估计方程特征筛选方法的思想建立特征筛选准则进行筛选,最后当迭代算法收敛时得到最终的筛选子模型;引入随机模拟和酵母细胞周期循环基因表达数据集对迭代的动态特征筛选方法和基于广义估计方程的特征筛选方法以及其他 2 个经典的独立特征筛选方法进行测试,结果表明:迭代的动态特征筛选方法不仅可以快速地筛选出重要协变量,而且还能够更加灵活地处理纵向数据的组内相关性,拥有更高的筛选精度。 |
| 关键词: 超高维纵向数据 特征筛选 修正的Cholesky分解 广义估计方程 动态协方差建模 |
| DOI： |
| 分类号: |
| 基金项目: |
|
| Feature Selection for Longitudinal Data Based on Dynamic Covariance Modeling |
|
CHEN Xinyue
|
|
School of Mathematics and Statistics, Southwest University, Chongqing 400715, China
|
| Abstract: |
| In order to make statistical analysis effective feature screening has been widely studied by many scholars in the ultra-high dimensional field. Aiming at the problem that the existing feature screening methods cannot flexibly deal with the intra-group correlation of ultra-high dimensional longitudinal data an iterative feature screening method based on dynamic covariance modeling was proposed. This method is called the iterative dynamic feature screening method. At each iteration the modified Cholesky decomposition was used to replace the static covariance matrix modeling method to dynamically model the intra-group covariance matrices of longitudinal data to obtain the flexible estimators of them and then these estimators were substituted into the generalized estimating equation GEE to establish the feature screening
criteria for screening according to the idea of GEE-based screening procedure GEES . Finally the final submodel was obtained when the iterative algorithm converged. Random simulations and yeast cell-cycle gene expression dataset were introduced to test the iterative dynamic feature screening method GEES and the other two classical independent feature screening methods. The results show that the iterative dynamic feature screening method can quickly screen out important covariates can deal with the intra-group correlation of longitudinal data more flexibly and has higher screening accuracy. |
| Key words: ultra-high dimensional longitudinal data feature screening modified Cholesky decomposition generalized estimating equations dynamic covariance modelling |
