Skip to content

Latest commit

 

History

History
140 lines (107 loc) · 4.99 KB

preparing_data.md

File metadata and controls

140 lines (107 loc) · 4.99 KB
Raw

Preparing Data

Splitting data

MLJ has two tools for splitting data. To split data vertically (that is, to split by observations) use partition. This is commonly applied to a vector of observation indices, but can also be applied to datasets themselves, provided they are vectors, matrices or tables.

To split tabular data horizontally (i.e., break up a table based on feature names) use unpack.

MLJBase.partition
MLJBase.unpack

Bridging the gap between data type and model requirements

As outlined in Getting Started, it is important that the scientific type of data matches the requirements of the model of interest. For example, while the majority of supervised learning models require input features to be Continuous, newcomers to MLJ are sometimes surprised at the disappointing results of [model queries](@ref model_search) such as this one:

using MLJ

X = (height   = [185, 153, 163, 114, 180],
     time     = [2.3, 4.5, 4.2, 1.8, 7.1],
     mark     = ["D", "A", "C", "B", "A"],
     admitted = ["yes", "no", missing, "yes"]);
y = [12.4, 12.5, 12.0, 31.9, 43.0]
models(matching(X, y))

Or are unsure about the source of the following warning:

Tree = @load DecisionTreeRegressor pkg=DecisionTree verbosity=0
tree = Tree();
julia> machine(tree, X, y)

julia> machine(tree, X, y)
┌ Warning: The scitype of `X`, in `machine(model, X, ...)` is incompatible with `model=DecisionTreeRegressor @378`:                                                                
│ scitype(X) = Table{Union{AbstractVector{Continuous}, AbstractVector{Count}, AbstractVector{Textual}, AbstractVector{Union{Missing, Textual}}}}
│ input_scitype(model) = Table{var"#s46"} where var"#s46"<:Union{AbstractVector{var"#s9"} where var"#s9"<:Continuous, AbstractVector{var"#s9"} where var"#s9"<:Count, AbstractVector{var"#s9"} where var"#s9"<:OrderedFactor}.
└ @ MLJBase ~/Dropbox/Julia7/MLJ/MLJBase/src/machines.jl:103
Machine{DecisionTreeRegressor,…} @198 trained 0 times; caches data
  args: 
    1:  Source @628 ⏎ `Table{Union{AbstractVector{Continuous}, AbstractVector{Count}, AbstractVector{Textual}, AbstractVector{Union{Missing, Textual}}}}`
    2:  Source @544 ⏎ `AbstractVector{Continuous}`

The meaning of the warning is:

  • The input X is a table with column scitypes Continuous, Count, and Textual and Union{Missing, Textual}, which can also see by inspecting the schema:
schema(X)
  • The model requires a table whose column element scitypes subtype Continuous, an incompatibility.

Common data preprocessing workflows

There are two tools for addressing data-model type mismatches like the above, with links to further documentation given below:

Scientific type coercion: We coerce machine types to obtain the intended scientific interpretation. If height in the above example is intended to be Continuous, mark is supposed to be OrderedFactor, and admitted a (binary) Multiclass, then we can do

X_coerced = coerce(X, :height=>Continuous, :mark=>OrderedFactor, :admitted=>Multiclass);
schema(X_coerced)

Data transformations: We carry out conventional data transformations, such as missing value imputation and feature encoding:

imputer = FillImputer()
mach = machine(imputer, X_coerced) |> fit!
X_imputed = transform(mach, X_coerced);
schema(X_imputed)

encoder = ContinuousEncoder()
mach = machine(encoder, X_imputed) |> fit!
X_encoded = transform(mach, X_imputed)

schema(X_encoded)

Such transformations can also be combined in a pipeline; see Linear Pipelines.

Scientific type coercion

Scientific type coercion is documented in detail at ScientificTypesBase.jl. See also the tutorial at the this MLJ Workshop (specifically, here) and this Data Science in Julia tutorial.

Also relevant is the section, Working with Categorical Data.

Data transformation

MLJ's Built-in transformers are documented at Transformers and Other Unsupervised Models. The most relevant in the present context are: ContinuousEncoder, OneHotEncoder, FeatureSelector and FillImputer. A Gaussian mixture models imputer is provided by BetaML, which can be loaded with

MissingImputator = @load MissingImputator pkg=BetaML

This MLJ Workshop, and the "End-to-end examples" in Data Science in Julia tutorials give further illustrations of data preprocessing in MLJ.