Merge pull request #39 from JuliaAI/dev

For a 0.2.6 release

Author: ablaom

Project.toml

@@ -1,7 +1,7 @@
 name = "FeatureSelection"
 uuid = "33837fe5-dbff-4c9e-8c2f-c5612fe2b8b6"
 authors = ["Anthony D. Blaom <anthony.blaom@gmail.com>", "Samuel Okon <okonsamuel50@gmail.com"]
-version = "0.2.5"
+version = "0.2.6"
 
 [deps]
 MLJModelInterface = "e80e1ace-859a-464e-9ed9-23947d8ae3ea"

src/models/rfe.jl

@@ -95,6 +95,12 @@ Train the machine using `fit!(mach, rows=...)`.
   `transform(mach, X)` above and predict using the fitted base model on the transformed
   table.
 
+!!! note
+
+    Because models wrapped in `RecursiveFeatureElimanation` are `Supervised`, the output
+    of `predict` is propagated in MLJ pipelines. To make this `transform` instead,
+    additionally wrap in `Transformer` as shown in the example below.
+
 # Fitted parameters
 
 The fields of `fitted_params(mach)` are:
@@ -146,6 +152,17 @@ predict(mach, Xnew)
 # transform data with all features to the reduced feature set:
 transform(mach, Xnew)
 ```
+
+To use `selector` as a transformer in an MLJ pipeline, you must explicitly wrap in
+`Transformer`, for otherwise it is the output of `predict` and not `transform` that is
+propagated to the next model in the pipeline:
+
+```julia
+pipe = Transformer(selector) |> ConstantRegressor()
+mach = machine(pipe, X, y) |> fit!
+predict(mach, Xnew) # prediction of `ConstantRegressor()` based on reduced features.
+```
+
 """
 function RecursiveFeatureElimination(
     args...;