How can I use knnimpute while having all rows of the input matrix with at least one missing value?

While trying to use knnimpute to fill in missing data, I get the following error. "All rows in the input data contain missing values. Unable to impute missing values."

It is not practical in most cases to have a feature (row in knnimpute data matrix argument) with no missing value. In the example above I would think given there are sufficient number of observations (columns) with complete values for each feature, this shouldn't cause any hiccup.

[Larali]

Hi Hambisa,

did you find any solution to this problem? Would be great to know, how you dealt with this issue.

Bests, Lara

[kokila Mani]

HI, You can transpose the matrix and try once again. (i.e)At=A'; knnimpute(At);

Answer 1

[Tim DeFreitas]

This is an older question, but in case anyone comes across this answer looking for further explanation:

knnimpute only calculates distance between observation columns using rows that do not contain NaN values. This is because if NaN rows were included, the distance between columns containing NaN values would also be NaN, and there would be no way to rank the k nearest neighbors for any observation.

You could force knnimpute to replace NaN values with the average of a feature across all non-NaN obseravations by adding a "feature" row, where each observation is identical, and then removing it:

A = [ 1 2 3 4 5 7 8 NaN; 8 7 6 5 4 3 2 NaN 1; 6 5 4 3 2 1 NaN 8 7];
A(4,:) = ones(1,9);

impA = knnimpute(A)
impA(4,:) = []

[Tsehaye Gebreteklie]

Thanks