Package name

[ericphanson]

@palday @kimlaberinto any bikeshedding on the naming? Scope is to provide exact or approximate solutions to combinatorial problems around partitioning objects, with an eye to usage in cross validation and test/train splits in ML tasks. Things like:

• I have N pieces of data grouped into M buckets. I know data within a bucket is correlated, so I don't want to split a bucket across CV splits or test/train splits. How do I partition my data into K equal groups? (That's the problem currently solved)
• My data is also labelled by class, and I want to obtain groups with equal class balance
• Instead of equal sized groups, I want an 80/20 split for test/train
• ...generalizations thereof, maybe multiple types of correlation, maybe more things to balance over, etc

name ideas:

• PartitioningProblems.jl
• DataPartitions.jl
• Partitioning.jl

Related work:

• https://github.com/JuliaML/MLDataPattern.jl
• https://alan-turing-institute.github.io/MLJ.jl/dev/evaluating_model_performance/#MLJBase.StratifiedCV
• https://github.com/kahypar/KaHyPar.jl ?
• https://github.com/JuliaSparse/Metis.jl ?

[ericphanson]

@palday said:

I don't like DataPartitions for some reason. Are there other relevant uses of "partition" that we don't want to clobber by taking this name?

Also, soon we'll have KnapsackProblems.jl.

[ericphanson]

Are there other relevant uses of "partition" that we don't want to clobber by taking this name?

Hm, yeah. I think Partitions.jl or Partitioning.jl might be a bit too grand, esp for a package in General where there could be bigger projects wanting that name.

There's graph partitioning which I think is actually related but formulated differently.

DataPartitions might be a little too scoped to ML stuff; I think it's kind of nice if this package is written a bit generically, since this kind of problem comes up in other contexts too, and we might get contributors from a broader set of fields that way.

[palday]

IntegerSplits?

[jrevels]

Is the existing name inaccurate?

assuming it is accurate, I kinda like the existing name (because it's precise/googleable), even if it's long

[kimlaberinto]

@jrevels I personally would say that the existing name is accurate. In fact it is nodding to the problem of the same name https://en.wikipedia.org/wiki/Multiway_number_partitioning

This package finds a solution to this problem, and is currently being applied internally at Beacon to find a nice configuration for a problem we face that is an instance of the Multiway Number Partitioning problem.

[ericphanson]

The context for the renaming is that we thought we might want to expand the scope beyond strictly the multiway partitioning problem-- that problem is the first bullet of https://github.com/beacon-biosignals/MultiwayNumberPartitioning.jl/issues/2#issue-972895738, but the others are kinda generalizations of it.

edit: but that doesn't mean we have to rename it, it could be implicitly MultiwayNumberPartitioning + generalizations thereof. Just wanted to say why we were talking about renaming it.

[ericphanson]

Ok let's just keep the name 😄