Open jolars opened 4 years ago
For me, choice between absolute errors (implemented now) and relative errors (described above) seems enough to cover a wide range of scenarios. One could consider other norms than euclidean, but that would rather seem to be a technical variation without much added practical value. Well, maybe max norm (\max(...) instead of \sum((...)^2).
Possible loss metrics include:
please comment if there are any other loss metrics that you would want to include