Fig. 2

The results of low-rank prediction in 119 datasets containing a combined total of 10,024 microarray slides at 750 iterations. Boxplots representing Frobenius relative error (top left) and spectral relative error (bottom left) in prediction of converged datasets, and the fraction of values known prior to prediction were varied. Edges of box represent 25 % and 75 % coverage, and the whiskers extend it to 99.73 % coverage, where outliers represent matrices generated using 10 datasets. Variation of omega relative error with the observability of three example datasets with a low Frobenius error (top right) and high Frobenius error (bottom right). Datasets with a high relative error in prediction (bottom right) have a corresponding high omega relative error