Semi-supervised imputation for microarray missing value estimation

Data missing is a kind of inevitable phenomenon in gene expression microarray experiments due to many factors. The integrity of the data plays a key role in the performance of the downstream analysis. Therefore, many developments have been achieved in the research on estimating missing values. However, when it comes to missing data with a large missing rate, most current estimation methods cannot obtain a high estimation precision. In this paper, induced by the thought of semi-supervised learning with collaborative training, we propose a new imputation method called COIM (COllaborative IMputation). COIM estimates missing values using collaborative imputation strategy based on Bayesian principal component analysis (BPCA) and local least squares (LLS). It exploits global correlation information and local structure in the missing dataset, by sharing the estimated results with each other between BPCA and LLS. Furthermore, COIM uses tactics of recovering genes that have less missing entries first. Numerical results demonstrate that COIM is superior to the comparative algorithms in terms of normalized root mean square error (NRMSE), especially for the datasets with large missing rates or less complete genes.