Predictions and Test & Score: MAPE produces value where it should fail

[wvdvegte]

What's wrong?
The formula for the Mean Absolute Percentage Error (see here or here) produces a zero division if the test data contains actual values that are zero.
Yet Predictions and Test & Score produce a (very large) value for MAPE if there are zeroes in the test data.

How can we reproduce the problem?
In the attached workflow, Formula is used to compute the Absolute Percentage Error for every record, of which MAPE is the mean over all rows. As it should, Formula fails if rows where the target equals zero are included, but at the same time Predictions and Test & Score produce very large values for MAPE.

MAPE 0div.zip

Perhaps my idea to include MAPE wasn't a very good one, since it is a problematic measure. Perhaps consider adding sMAPE (Python code provided here) or even replacing MAPE by it.

[janezd]

Scipy tries to be smart and divides by the smallest non-zero float, resulting in an essentially arbitrarily chosen large value.

MAPE is trivial to compute: I believe we should replace it with our own, proper implementation which returns inf (and fix the widget, if it can't handle it). This will make the score useless in case of zero values in test data, but it is what it is.

I like sMAPE. I think we should add it.