rlearn

Rlearn is an R package for a specific machine learning process. It provides variable selection and linear discriminant analysis model fitting and assessment which is linked to the variable selection results.

The variable selection routine produces a list of candidate models. The linear discriminant analysis (LDA) routine fits and assesses the performance of LDA models for each candidate model from variable selection using leave-one-out cross validation. Finally, a function is provided to fit an LDA on an entire dataset for one model in the variable selection candidate list.

Components

Variable Selection

Given a dataset, identify the subsets which can best represent the information in the entire dataset.

Wraps the subselect library https://cran.r-project.org/web/packages/subselect/subselect.pdf which assesses ability of data subsets to represent entire datasets and finds optimal subsets under various constraints

Computes total and between-group matrices of Sums of Squares and Cross-Product (SSCP) deviations in linear discriminant analysis. http://www.ibm.com/support/knowledgecenter/SSLVMB_21.0.0/com.ibm.spss.statistics.help/alg_discriminant_basicstats_w.htm http://sites.stat.psu.edu/~ajw13/stat505/fa06/12_1wMANOVA/03_1wMANOVA_multi.html http://stats.stackexchange.com/questions/82959/how-is-manova-related-to-lda http://stats.stackexchange.com/questions/48786/algebra-of-lda-fisher-discrimination-power-of-a-variable-and-linear-discriminan#48859

The SSCP matrix is used to find subsets of variables which best represents the information in the entire data set given a criterion (score function) and constrains (number of variables in the subset). The Multivariate Linear Chi-squared / Pillai-Bartlett trace is used to score the variable subsets using class prediction rates and observed class frequencies. https://en.wikipedia.org/wiki/Contingency_table#Squared_normal_distributions_in_contingency_tables

Linear Discriminant Analysis

Given variable subsets as chosen by variable selection, calculate discriminant functions and classification scores for each variable subset. Jacknife/Leave-one-out cross validation is first used to calculate cross tabulations (for confusion matrices) and error rates. Finally, LDA is run on all observations, for each model, to get discriminant functions and between/within group variance ratios for each model.

Data processing

Data can be filtered on 1 column, with 1 value excluded. Degenerate variables (standard deviation of 0) are removed from input data before modeling. Variables which should be considered are included in a configuration file. See tests/data/expected for examples.

Installing

There are a couple installation options

• clone and install with R CMD
• use devtools to instrall directly from github
• install into a docker container using the provided Dockerfile

Docker is the suggested method, as it provides a more consistent environment.

R CMD

$ git clone git@github.com:tesera/rlearn.git
$ cd rlearn
$ R CMD BUILD .
$ R CMD INSTALL rlearn_1.0.0.tar.gz

Devtools

$ R
> library('devtools')
> install_github(repo="tesera/rlearn", ref="master", auth_token="<your_github_personal_access_token>")

Docker

Clone the repository and build the docker image

$ git clone git@github.com:tesera/rlearn.git
$ cd rlearn
$ docker built -t rlearn .

Grab a coffee, building the image will take a few minutes

Step 1 : FROM r-base:latest
latest: Pulling from library/r-base
9cd73496e13f: Downloading [=============================>                     ] 24.73 MB/42.07 MB
f10af350cd29: Download complete
eea7b33eea97: Download complete
c91475e50472: Download complete
1e5e5f6785b4: Download complete
8c4091261ff6: Downloading [>                                                  ] 5.919 MB/322.1 MB
...

To test if the image built successfully, run the following command

docker run -it rlearn

That should drop you into an interactive R session where you can import and use rlearn

> library(rlearn)
> 

Usage

Start R

$ R
# or, if you are using docker
$ 

Development

Setup

All contributors are welcome! To get started developing on rlearn you will need docker

One you are setup with docker, clone this repo

$ git clone git@github.com:tesera/rlearn.git

Enter the top level directory

cd rlearn
docker-compose run dev

Now you are in the docker container - install the dependencies required for rlearn

r install-dependencies.r

And you're all set to make changes to rlearn!

To use your active development version of rlearn start an R session in the docker container

docker-compose run dev R

And load all the package resources

library(devtools)
load_all('.')

Testing

Unit tests are written using the testthat package.

Unit tests are required for new functionality. Changes to existing codebase should not break existing tests, or existing tests should be updated if appropriate.

Run the tests as follows

docker-compose run dev
$ r ./test.r

Documentation

If your changes affect documentation, rebuild the documentation as follows:

$ docker-compose run dev
$ R
> library(devtools)
> document()

Contribution Guidelines

• If you would like to contribute changes to rlearn, please follow this guide to fork, clone, create a branch, make your changes, push your branch to your fork, and open a pull request. Don't forget to run the tests!
• Please follow the Wickham Style for your contributions to rlearn.

Getting Help

For assistance with usage or development of rlearn, please file an issue on the issue tracker.