v0.18.20

MLJBase v0.18.20

Diff since v0.18.19

• (enhancement) Integrate new outlier detection types added to MLJModelInterface (JuliaAI/MLJ.jl#780, JuliaAI/MLJModelInterface.jl#113) and adapt performance evaluation methods to handle them (#628)

Relating to measures:

• (enhancement) Measures: Add SphericalScore and LogScore (negative of LogLoss).
• (enhancement) Extend the proper scoring rules SphericalScore, LogScore and BrierScore to handle Continuous and Count data. Supported distributions types from Distributions are: Uniform, Normal, Exponential, Logistic, Chi, Chisq, Beta, Gamma, Cauchy, Poisson, DiscreteUniform, DiscreteNonParameteric (#627)
• (enhancement) Add missing and NaN support for all measures, excluding AreaUnderCurve and measures from LossFunctions.jl (which imported library does not support) (#616)
• (enhancement) Add skipinvalid(y) and skipinvalid(yhat, y) methods. The first returns an iterator that skips missing and NaN values - similar to skipmissing and is performant. The second returns the flattened forms of yhat and y with invalid entries removed "commensurately" from both yhat and y, meaning an element of either argument is skipped even if valid, if the corresponding element of the other argument is invalid (#627). This method is necessarily less efficient and provided for convenience for pre-processing data for the measures which do not support invalid entries (#627, #616)
• (enhancement) Allow most measures for Finite data to be called with "raw" data, that is, data that is not wrapped as CategoricalArray. This includes ConfusionMatrix. A warning is issued to indicate order ambiguity, with the usual suggestion to coerce to OrderedFactor to suppress the warning (#627)
• (enhancement) Allow measures to be called on arrays, and not just vectors (#627) but see remaining limitations at #631.
• (API) Make implementing new measures simpler (#627) and less error-prone. See this guide for details.
• (enhancement) Introduce new method MLJBase.call(measure, args...) to call a measure without applying dimension or pool checks.
• (bug fix) Prevent weights passed to measures from Loss functions.jl from being normalized (#626)

Closed issues:

• Add BrierScore for Continuous targets and assorted Distributions (#395)
• Add support for missing/NaN values where possible in measures (#616)
• Improve the doc-string for unpack (#620)
• Skip NaN in aggregation of measures (#622)
• Weights passed to measures from LossFunctions are being normalized (#626)

Merged pull requests:

• Add Continuous and Count versions of Brier score (#623) (@ablaom)
• Skip NaN as well as missing in measure aggregation (#624) (@ablaom)
• Add Infinite versions of the proper scoring rules spherical, log (#625) (@ablaom)
• Big measures cleanup (#627) (@ablaom)
• Outlier Detection Integration (#628) (@davnn)
• For a 0.18.20 release (#632) (@ablaom)
• Last minute bump of Distributions compat (#633) (@ablaom)