Add method to facilitate tabulation of multiple performance evaluations.

[ablaom]

This PR is a follow up for the enhancements at #1034.

This PR extends DataAPI.describe (also extended by DataFrames) to summarise an MLJ performance evaluation, as a named tuple. Multiple evaluations can then be combined in a table:

using MLJ
X, y = @load_iris # a vector and a table

# instantiate two models:
knn = (@load KNNClassifier pkg=NearestNeighborModels)()
tree = (@load DecisionTreeClassifier pkg=DecisionTree)()

named_models = [
    "Dummy" => ConstantClassifier(),  # a built-in model
    "K-nearest neighbors" => knn,
    "Decision Tree" => tree,
]
performance_evaluations = evaluate(named_models, X, y; measures=[brier_score, accuracy])
julia> describe(performance_evaluations[1])
(tag = "Dummy", BrierScore = -0.573 ± 0.1, Accuracy = 0.33 ± 0.23)

table = describe.(performance_evaluations)
julia> pretty(table)
┌─────────────────────┬──────────────────────┬──────────────────────┐
│ tag                 │ BrierScore           │ Accuracy             │
│ String              │ Measurement{Float64} │ Measurement{Float64} │
│ Textual             │ Continuous           │ Continuous           │
├─────────────────────┼──────────────────────┼──────────────────────┤
│ Dummy               │ -0.573±0.1           │ 0.33±0.23            │
│ K-nearest neighbors │ -0.21±0.21           │ 0.92±0.18            │
│ Decision Tree       │ -0.00118977±0.0      │ 1.0±0.0              │
└─────────────────────┴──────────────────────┴──────────────────────┘

I'm not wedded to the name describe. Maybe better is summarize but I don't know other examples of summarize in the ML ecosystem and I was reluctant to add yet another method to the namespace. Thoughts or suggestions welcome.

cc @OkonSamuel @LucasMatSP

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 89.08%. Comparing base (be6d92f) to head (24d7d65).

Additional details and impacted files

@@            Coverage Diff             @@
##              dev    #1045      +/-   ##
==========================================
- Coverage   90.26%   89.08%   -1.19%     
==========================================
  Files          33       34       +1     
  Lines        2600     2574      -26     
==========================================
- Hits         2347     2293      -54     
- Misses        253      281      +28     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[tylerjthomas9]

The describe name is more intuitive and consistent with other similar naming patterns I've seen.

One small thing I noticed is that individualize iterates over the input twice. In practice, I don’t think it will cause issues here because nobody would feed in an IO stream or other non-re-iterable collection, but the docstring could be tweaked to reflect this

Overall, it looks great to me.

[LucasMatSP]

Personally, I prefer summarize, but I don't have a strong opinion about this. The feature is nice

[ablaom]

Thanks for the input! And thank you kindly @tylerjthomas9 for the review.

One small thing I noticed is that individualize iterates over the input twice.

Good catch. I have tweaked the docstring. But I also realised we already have a similar helper method in the pipeline code, called individuate. Therefore I will refactor this and move the method to "src/utilities.jl".

Regarding the name, I'm inclining towards describe, as I couldn't find summarize and Base.summary doesn't apply. But two other alternatives I thought of are:

• overload NameTuple(::AbstractPerformanceEvaluation)
• report, which we already have for machines

Does anyone like these better than describe?

[tylerjthomas9]

Regarding the name, I'm inclining towards describe, as I couldn't find summarize and Base.summary doesn't apply. But two other alternatives I thought of are:

• overload NameTuple(::AbstractPerformanceEvaluation)

• report, which we already have for machines

Does anyone like these better than describe?

I am a fan of describe, but I may be alone on this one. It's intuitive to me and fits with other libraries like DataFrames.jl's describe.