I have a use-case for NMF which fits the W1, H matrices using one dataset, then solves for W2 (using the same H) for a second dataset. The point is to use a reference data set to produce the "components" and then fit a target data set to those components.
In sklearn, this is achievable by calling fit on the reference data and then transform on the target data. This makes a lower-level call to non_negative_factorization which has an option update_H::Bool.
This seems like something that could be added as an option to the solve! function.
I have a use-case for NMF which fits the
W1, H
matrices using one dataset, then solves forW2
(using the sameH
) for a second dataset. The point is to use a reference data set to produce the "components" and then fit a target data set to those components.In sklearn, this is achievable by calling
fit
on the reference data and thentransform
on the target data. This makes a lower-level call tonon_negative_factorization
which has an optionupdate_H::Bool
.This seems like something that could be added as an option to the
solve!
function.