Triple imputation for microarray missing value estimation

Data obtained from gene expression microarray experiments always suffer from missing values due to various reasons. However, complete gene expression data are of great importance to many gene expression data analysis issues. Therefore, imputation methods with high estimation precision are critical to further data analysis. In this paper, inspired by the idea of semi-supervised learning with tri-training, we propose a novel imputation method called TRIIM (TRIple IMputation). TRIIM estimates missing values using triple imputation strategies based on Bayesian principal component analysis (BPCA), local least squares (LLS) and expectation maximization (EM). The data properties of global correlation information, local structure and data distribution are all considered properly. It is implemented by sharing the estimated values of any two algorithms' cooperation to the rest at each step, and assembling combinations of all imputation results finally. Experimental results on four real microarray matrices demonstrate that TRIIM achieves better performance than the comparative algorithms in terms of normalized root mean square error (NRMSE), even in the case of microarray dataset with large missing rates and few complete genes.