median absolute error

[oxinabox]

I found a paper using this

Its simple enough to implement:

median_absolute_error(ŷ, y) =median(abs.(ŷ .- y))
const mdae = median_absolute_error

Its not the fastest measure to compute, but on most sizes of test data it is going to be cheap vs training.

[ablaom]

PR welcome.

This probably follows the pattern of RSquared and would have the same traits.

You will want to add missing handling, which you can do by including the supports_missings_measure wrapper in this pipeline.

You can add the aliases median_absolute_error and mdae when you "register" the new constructor.