`cut` into `Interval`s

[jariji]

Currently cut returns String-valued categories but Interval-valued categories from IntervalSets.jl could be nicer. What do you think about this?

[bkamins]

What we would need to do is to allow types from IntervalSets.jl to be category levels (now only char, string and number are allowed). But I think it would make sense to allow them. IntervalSets.jl is a relatively light package.

@nalimilan - what do you think?

[nalimilan]

I'm not sure it would really be useful. Strings are convenient and flexible, and cut is intended to allow naming classes like "20-24 years", "Q1" or "Low". Do you have a concrete use case in mind?

IntervalSets could provide another function to cut a numeric variable into proper intervals. If you have intervals, you probably don't need a CategoricalArray since intervals are naturally sorted in the correct order (contrary to string classes).

Also in terms of implementation, allowing Interval values in CategoricalArrays would require taking a dependency on IntervalSets. I don't think an extension would work unfortunately. It's annoying we have to limit supported types like this but it's required to limit invalidations...

[jariji]

Do you have a concrete use case in mind?

I want to be able to use the IntervalSets.jl functions. At the moment I am just using Interval manually without CategoricalArrays, which is fine, but I am

• extracting the leftendpoint and rightendpoint
• checking which bin a new value belongs in, using findfirst(∋(x), intervals)

[nalimilan]

That's not what I call a concrete use case. ;-) What kind of data are you processing? For what goal?

[bkamins]

I assume that a basic use case is that:

1. originally, you have a continuous variable.
2. you bin it (e.g. for an input into some model)
3. later you get new data (continuous), and you would want to learn into what level of the categorical (binned) variable it would fall.

Of course, all this can be done without CategoricalArrays.jl, but I assume that @jariji wants some convenience and consistency here.

[jariji]

I have duration data and I'm partitioning it into intervals for a model where each interval has a parameter, something like 90-day-period fixed effects.

Btw I'm using numbers for time durations but could imagine using Dates.Days or something.

IntervalSets could provide another function to cut a numeric variable into proper intervals.

This might be the best solution, though I guess IntervalSets.jl doesn't have the concept of a partition, so it doesn't know that some interval v_1 is part of a partition v of intervals dividing a space.

[nalimilan]

I see. Do you then get a single new value to fit into existing intervals? Or a vector of new values?

[jariji]

I'm just doing it on the array, so I don't need any special "categorize a new value" function -- I'm basically just implementing cut with findfirst. But I wouldn't put too much weight on whatever I happen to be doing at the moment.