Comparison of machine learning algorithms on missing values estimation accuracy of microarray datasets

Existence of missing values in DNA microarray data would decrease the accuracy of regulatory gene networks construction and may cause mistake in clustering and classifying gene expression for downstream analysis. Therefore, missing value imputation is a pivotal step in preprocessing of DNA microarray data. Selection a proper algorithm for achieving the most accurate conclusions in missing values imputation remains to be quite compelling. In this study, three microarray datasets were used to compare the performance of different machine learning algorithms in imputing DNA microarray missing values. In this way, after determining the dimensions of matrix of expression data and normalizing the data, different missing percentages were applied on each datasets. By running 11 machine learning algorithms on these datasets, the accuracy of each algorithm under different conditions were measured. Based on the results, the accuracy of different algorithms depended on missing value percentages and its distribution in the dataset. Also, the number of experimental samples in the datasets affected the accuracy of missing values imputation algorithms. The results showed a decreasing trend in accuracy by increasing the percentage of missing data in the dataset. In general, Least Square Adaptive and Local Least Square algorithms shown to be more robust in terms of accuracy when the level of missing values percentage increased in the dataset. Therefore, we would suggest these algorithms could be considered in working out sound missing values imputation in DNA microarray data.