`predict_spatial()`: Create unsupervised task internally?

[mlr-org]

`predict_spatial()`: Create unsupervised task internally?

[pat-s]

@be-marc This got a few questions during the last workshop. I wonder if we should handle this internally?
I can't remember if something was preventing is from doing so?

[be-marc]

No, not technically. I'm not sure to what degree we should hide things from the user.

[pat-s]

Think of the average (spatial) modeling user: the user is doing a prediction and then sees the use of as_unsupervised_task() and thinks: "wait, what does this do now and why is it needed?". The point is that usually people rarely use unsupervised tasks and there is no other occurrence AFAICS where this is needed to execute a prediction scenario.

It then happened that I had to explain this (or I rather said: "well, there are technical reasons why we need to do so but you don't need to worry" because there is no time in a presentation to go deeper into this and it might confuse people even (more) at this point).

I think it might be better if

• we allow passing the raster directly
• document the use of of the unsupervised task in the help page of predict_spatial()

In the end users want to predict on a raster and they don't really care if this one needs to be wrapped into an unsupervised task, mainly because for internal mlr3 reasons.

What do you think?

[be-marc]

I think this only works if the user is not making a mistake. For example, if the user passes a raster to predict_spatial() with a missing spectral band. He will get an error from the learner that TaskUnsupervised misses a feature. The user did not create a TaskUnsupervised so this might be confusing.

We often pass a task to a learner for making predictions. It's one of the first things the user learns in the book. I think the API is fine like this.

[be-marc]

It then happened that I had to explain this (or I rather said: "well, there are technical reasons why we need to do so but you don't need to worry" because there is no time in a presentation to go deeper into this and it might confuse people even (more) at this point).

We could just say that in mlr3 predictions are usually made on tasks.

[be-marc]

I changed my mind. Passing spatial objects is now supported. TaskUnsupervised still works but will be removed from the docs.

#82