Figure 1

WeNNI is the most accurate imputation method. Performance of the five imputation methods with varying β. As explained in the Methods section, larger β changes weights to smaller values. In non-weighted methods β is the SNR cutoff. The increase in MSD for large β is an effect from too many missing values, which implies imputation breaks down. The standard error of means are within the line thicknesses. (A) Breast cancer data. WeNNI (black line) has the lowest MSD and the weighted methods perform better than the non-weighted methods. All methods have a minimum MSD around β = 0.2. (B) Melanoma data. WeNNI (black line) has the lowest MSD and the weighted methods perform better than the non-weighted methods. All methods have a minimum MSD around β = 0.6. (C) Mycorrhiza data. WeNNI (black line) retains the lowest MSD, whereas KNNimpute (red line) performs better that the weighted reporter average method. This may be explained as an effect of a different experimental design as discussed in the text. The minimum MSD is found in a β range 0.3–1 for the different methods.