Performance evaluation of L<inf>1</inf>-norm-based microarray missing value imputation

Abstract

l 1 -norm minimization was utilized in the imputation of microarray missing values, which is an important procedure in bioinformatics experiments. Two l 1 approaches, based on the framework of local least squares (LLS) and iterative biclusterbased least squares (bicluster-iLLS) respectively, were employed. Imputed datasets of the l 1 approaches were compared with those of traditional l 2 methods. The imputation error rates showed that the assumption of sparsity is not supported by the microarray datasets. Singular value decompositions in biclusters and in the neighborhoods of target gens were computed to show the structure of a microarray dataset. The coefficients of l 1 minimization solutions were also analyzed to reveal possible reasons for the performance of l 1 approaches.