Add PolynomialTransformer

[ablaom]

This PR addresses #45 . InteractionTransformer(; kwargs...) is now equivalent to* PolynomialTransformer(; interactions_only=true, kwargs...) but using InteractionTranformer triggers a depwarn.

I've renamed the field order to degree (seems more usual for a polynomial) but order is retained as a kwarg, aliased to degree.

Here's an example drawn from the new doc string:

using MLJ

X = (
    A = [1, 2, 3],
    B = [4, 5, 6],
    C = [7, 8, 9],
    D = ["cat", "dog", "rat"]
)

transformer = PolynomialTransformer(degree=2, features=[:A, :B])
mach = machine(transformer)
julia> transform(mach, X) |> pretty
┌───────┬───────┬───────┬─────────┬───────┬───────┬───────┐
│ A     │ B     │ C     │ D       │ A_A   │ A_B   │ B_B   │
│ Int64 │ Int64 │ Int64 │ String  │ Int64 │ Int64 │ Int64 │
│ Count │ Count │ Count │ Textual │ Count │ Count │ Count │
├───────┼───────┼───────┼─────────┼───────┼───────┼───────┤
│ 1     │ 4     │ 7     │ cat     │ 1     │ 4     │ 16    │
│ 2     │ 5     │ 8     │ dog     │ 4     │ 10    │ 25    │
│ 3     │ 6     │ 9     │ rat     │ 9     │ 18    │ 36    │
└───────┴───────┴───────┴─────────┴───────┴───────┴───────┘

transformer = PolynomialTransformer(degree=3, interactions_only=true)
mach = machine(transformer)
julia> transform(mach, X) |> pretty
┌───────┬───────┬───────┬─────────┬───────┬───────┬───────┬───────┐
│ A     │ B     │ C     │ D       │ A_B   │ A_C   │ B_C   │ A_B_C │
│ Int64 │ Int64 │ Int64 │ String  │ Int64 │ Int64 │ Int64 │ Int64 │
│ Count │ Count │ Count │ Textual │ Count │ Count │ Count │ Count │
├───────┼───────┼───────┼─────────┼───────┼───────┼───────┼───────┤
│ 1     │ 4     │ 7     │ cat     │ 4     │ 7     │ 28    │ 28    │
│ 2     │ 5     │ 8     │ dog     │ 10    │ 16    │ 40    │ 80    │
│ 3     │ 6     │ 9     │ rat     │ 18    │ 27    │ 54    │ 162   │
└───────┴───────┴───────┴─────────┴───────┴───────┴───────┴───────┘

An additional change: Somewhat confusingly, we had and an empty src/utils.jl but test/utils.jl defines only functions needed to run tests (dataset generators). I moved some code shared by both InteractionTransformer and PolynomialTransformer to src/utils.jl and moved the contents of test/utils.jl to a new file test/test_utils.jl so I could add actual tests of src/utils.jl in test/utils.jl.

*This is not quite true. InteractionTransformer always returns a NamedTuple of Vectors. PolynomialTransformer tries to return a table of the same type as the input table, using Tables.materalizer(input_table).

[ablaom]

On another note, the logic for actually generating column combinations in InteractionTransformers uses Combinatorics.jl. I couldn't see similar functionality there to generate complete lists of monomials so had to add custom code for that. It's possible the new code could also be used to get rid of Combinatorics.jl altogether, but I haven't done that here.

[ablaom]

@olivierlabayle Are you willing and able to provide a review here, in the next 3-4 weeks, say? This PR makes a replacement for your InteractionTransformer, as set out in the referenced issue.

[olivierlabayle]

@ablaom Yes thank you, please assign me and I'll review!

[ablaom]

Thank you very much.

[olivierlabayle]

This looks good to me, two little caveats:

• You forgot to include the new test file in runtests.jl
• The "combination with replacement" exists in Combinatorics.jl. Since the MLJTransforms already depends on it, it might be worth outsourcing the algorithm to them.

[olivierlabayle]

why not throwing here?

[ablaom]

clean! mutates the model parameter so that it is valid, so no need to throw an exception. Or am I missing the point of your question?

[olivierlabayle]

I suppose I am wondering whether it would be better to explicitly throw if a user's code does not make sense rather than assuming a fallback to order=2?

[olivierlabayle]

You probably want to include the new include("transformers/other_transformers/polynomial_transformer.jl")

[ablaom]

Done in 4d58a8c

[olivierlabayle]

any reason to keep the MLJTransforms prefix here and line 142?

[ablaom]

Not necessary but helpful for maintenance: The method actualfeaures, while not defined in the current file, is defined in MLJTransforms somewhere, and not from the namespace of some other package. At least, this is helpful to me.

[olivierlabayle]

Your implementation seems to produce the same results but this is also defined in Combinatorics: Combinatorics.with_replacement_combinations or am I mistaken? At least the tests also pass with this function.

[ablaom]

Ah - I completely missed this, thank you. In eeab9bf I've swapped this in and dumped my orderedwords method.

[olivierlabayle]

Thank you @ablaom , all good for me, I'll let you make the call on the clean! vs throw as it is rather a design choice.

[ablaom]

@olivierlabayle Thanks for your prompt review. As indicated, I adopted all your suggestions.

Maintaining the quality of the MLJ ecosystem depends crucially on volunteer PR reviews such as this one. ❤️