mlrCPO: Composable Preprocessing Operators for mlr
GSoC 2017 Project: Operator Based Machine Learning Pipeline Construction
> task = iris.task
> task %<>>% cpoScale(scale = FALSE) %>>% cpoPca() %>>% # pca
> cpoFilterChiSquared(abs = 3) %>>% # filter
> cpoModelMatrix(~ 0 + .^2) # interactions
> head(getTaskData(task))
PC1 PC2 PC3 PC1:PC2 PC1:PC3 PC2:PC3 Species
1 -2.684126 -0.3193972 0.02791483 0.8573023 -0.07492690 -0.008915919 setosa
2 -2.714142 0.1770012 0.21046427 -0.4804064 -0.57122986 0.037252434 setosa
3 -2.888991 0.1449494 -0.01790026 -0.4187575 0.05171367 -0.002594632 setosa
4 -2.745343 0.3182990 -0.03155937 -0.8738398 0.08664130 -0.010045316 setosa
5 -2.728717 -0.3267545 -0.09007924 0.8916204 0.24580071 0.029433798 setosa
6 -2.280860 -0.7413304 -0.16867766 1.6908707 0.38473006 0.125045884 setosa
"Composable Preprocessing Operators" are an extension for the mlr ("Machine Learning in R") project which represent preprocessing operations (e.g. imputation or PCA) in the form of R objects. These CPO objects can be composed to form more complex operations, they can be applied to data sets, and they can be attached to mlr Learner objects to generate complex machine learning pipelines that perform both preprocessing and model fitting.
CPOs are created by calling a constructor.
> cpoScale()
scale(center = TRUE, scale = TRUE)
The created objects have Hyperparameters that can be manipulated using getHyperPars, setHyperPars etc, just like in mlr.
> getHyperPars(cpoScale())
$scale.center
[1] TRUE
$scale.scale
[1] TRUE
> setHyperPars(cpoScale(), scale.center = FALSE)
scale(center = FALSE, scale = TRUE)
The %>>%-operator can be used to create complex pipelines.
> cpoScale() %>>% cpoPca()
(scale >> pca)(scale.center = TRUE, scale.scale = TRUE)
This operator can also be used to apply an operation to a data set:
> head(iris %>>% cpoPca())
Species PC1 PC2 PC3 PC4
1 setosa -5.912747 2.302033 0.007401536 0.003087706
2 setosa -5.572482 1.971826 0.244592251 0.097552888
3 setosa -5.446977 2.095206 0.015029262 0.018013331
4 setosa -5.436459 1.870382 0.020504880 -0.078491501
5 setosa -5.875645 2.328290 -0.110338269 -0.060719326
6 setosa -6.477598 2.324650 -0.237202487 -0.021419633
Or to attach an operation to an MLR Learner, which extends the Learner's hyperparameters by the CPO's hyperparameters:
> cpoScale() %>>% makeLearner("classif.logreg")
Learner classif.logreg.scale from package stats
Type: classif
Name: ; Short name:
Class: CPOLearner
Properties: numerics,factors,prob,twoclass
Predict-Type: response
Hyperparameters: model=FALSE,scale.center=TRUE,scale.scale=TRUE
Get a list of all CPOs by calling listCPO().
Install mlrCPO from CRAN, or use the more recent GitHub version:
devtools::install_github("mlr-org/mlrCPO")
To effectively use mlrCPO, you should first familiarize yourself a little with mlr. There is an extensive tutorial online; for more resources on mlr, see the overview on mlr's GitHub page.
To get familiar with mlrCPO, it is recommended that you read the vignettes. For each vignette, there is also a compact version that has all the R output removed.
- First Steps: Introduction and short overview (compact version).
- mlrCPO Core: Description of general tools for
CPOhandling (compact version). - Builtin CPOs: Listing and description of all builtin
CPOs (compact version). - Custom CPOs: How to create your own
CPOs. (compact version). - CPO Internals: A small intro guide for developers into the code base. See the
infodirectory for pdf / html versions.
For more documentation of individual mlrCPO functions, use R's built-in help() functionality.
The foundation of mlrCPO is built and is reasonably stable, only small improvements and stability fixes are expected here. There are still many concrete implementations of preprocessing operators to be written.
mlrCPO is a free and open source software project that encourages participation and feedback. If you have any issues, questions, suggestions or feedback, please do not hesitate to open an "issue" about it on the GitHub page!
In case of problems / bugs, it is often helpful if you provide a "minimum working example" that showcases the behaviour (but don't worry about this if the bug is obvious).
Please understand that the resources of the project are limited: response may sometimes be delayed by a few days, and some suggestions may not not make it to become features for a while.
Pull Requests that fix small issues are very welcome, especially if they contain tests that check for the given issue. For larger contributions, or Pull Requests that add features, please note:
-
Adding new
CPOs is always welcome. Please have a look at a few examples in the current codebase (the PCA CPO and the corresponding tests file are good for this, and show that adding a CPO does not require a lot of code) to familiarise yourself with the conventions. ACPOthat comes with documentation, in particular also documenting theCPOTrainedstate, and with tests, is most likely to get merged quickly. -
Adding or changing features of the backend, or changing the functioning of the backend, is a more complicated story. If a Pull Request is incongruent with the "vision" behind
mlrCPO, or if it appears to put a large burden on themlrCPOdevelopers in the long term relative to the problems it solves, it may have a slim chance of getting merged. Therefore, if you plan to make a contribution changingCPOcore behaviour, it is best if you first open an "issue" about it for discussion.
When creating Pull Requests, please follow the Style Guide. Adherence to this is checked by the CI system (Travis). On Linux (and possibly Mac) you can check this locally on your computer using the quicklint tool in the tools directory. This is recommended to avoid frustrating failed builds caused by style violations.
Before merging a Pull Request, it is possible that an mlrCPO developer makes further changes to it, e.g. to harmonise it with conventions, or to incorporate other ideas.
When you make a Pull Request, it is assumed that you permit us (and are able to permit us) to incorporate the given code into the mlrCPO codebase as given, or with modifications, and distribute the result under the BSD 2-Clause License.
There are other projects that provide functionality similar to mlrCPO for other machine learning frameworks. The caret project provides some preprocessing functionality, though not as flexible as mlrCPO. dplyr has similar syntax and some overlapping functionality, but is focused ultimately more on (manual) data manipulation instead of (machine learning pipeline integrated) preprocessing. Much more close to mlrCPO's functionality is the Recipes package. scikit learn also has preprocessing functionality built in.
The BSD 2-Clause License