How to pool ROC curves in R to better understand a model's performance (CC135)

August 9, 2021 • PD Schloss • 2 min read • •

In this Code Club, Pat shows how he would pool ROC curves so that you can directly assess a model’s sensitivity for specificity. The area under the receiver operator characteristic (ROC) curve (AUC) is a useful metric of performance, but it isn’t always the best way to assess performance since it looks over all possible specificities. The challenge is that with the mikropml framework we get one ROC curve per 80/20 training-testing split and we need to pool the curves to get a composite ROC curve. Even if you don’t care about ROC curves, this episode is sure to have a lot of value for you including a little known R tip towards the end of the episode!

Pat will use functions from the mikropml R package and the ggplot2 and dplyr packages in RStudio.

Code

You can browse the state of the repository at the

Installations

If you haven’t been following along, you can get caught up by doing the following:

(windows) Install the Ubuntu Linux BASH shell for Windows 10

(mac) Install homebrew and git

  /usr/bin/ruby -e "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/master/install)"
  brew install git

To get to where we are at the beginning of this episode (you won’t have the same issue numbers at Pat)…
- Set up a GitHub account
- Create a new GitHub repository
  - Call it “mikropml_demo”
  - Make it Public
  - Don’t check the box next to “Initialize this repository with a README”
  - Click the green “Create repository” button
- Go to your command line and enter the following replacing <your_github_id> with your GitHub user id
```
git clone git@github.com:SchlossLab/mikropml_demo.git
cd mikropml_demo
git reset --hard 473f3da8502b49f1382e16826e6bcfc7b361b930
git remote set-url origin git@github.com:<your_github_id>/mikropml_demo.git
git push -u origin master
```
- Return to GitHub and refresh your browser.
- Go to the mikropml_demo directory on your computer and double click on the mikropml_demo.Rproj icon. This will launch RStudio and you’ll be good to go.