Additional loo utilities

[LeeviLindgren]

Hi!

This PR adds functions to compute additional LOO utilities: mean absolute error, mean squared error, and root mean squared error for regression problems and accuracy and balanced accuracy for binary classification. I have already discussed this with @avehtari and opened a PR for more discussion and feedback.

[codecov-commenter]

Codecov Report

Merging #202 (4df55e8) into master (0284928) will decrease coverage by 1.68%. The diff coverage is 100.00%.

:exclamation: Current head 4df55e8 differs from pull request most recent head ad74be2. Consider uploading reports for the commit ad74be2 to get more accurate results

@@            Coverage Diff             @@
##           master     #202      +/-   ##
==========================================
- Coverage   95.15%   93.47%   -1.69%     
==========================================
  Files          28       29       +1     
  Lines        2663     2712      +49     
==========================================
+ Hits         2534     2535       +1     
- Misses        129      177      +48     

Impacted Files	Coverage Δ
R/loo_predictive_error.R	100.00% <100.00%> (ø)
R/zzz.R	76.92% <0.00%> (-23.08%)	:arrow_down:
R/effective_sample_sizes.R	84.25% <0.00%> (-13.39%)	:arrow_down:
R/importance_sampling.R	84.04% <0.00%> (-8.52%)	:arrow_down:
R/loo.R	93.00% <0.00%> (-5.35%)	:arrow_down:
R/loo_moment_matching.R	97.63% <0.00%> (-1.58%)	:arrow_down:
R/loo_subsample.R	94.80% <0.00%> (-0.51%)	:arrow_down:

:mega: We’re building smart automated test selection to slash your CI/CD build times. Learn more

[avehtari]

Looks good! I hope @jgabry has also time to have a look? PR is small and clear, so I hope it's not taking much time

[avehtari]

I think this PR could be merged adding just the loo_predictive_error function, but it would be useful to think the bigger picture and how we could support loo_compare with different criteria

[jgabry]

Looking at this now. Sorry for the delay!

[jgabry]

@LeeviLindgren Thanks for the PR! Sorry for the delay in reviewing it. I'll make a few review comments now, but can you also edits from maintainers? You can do that by following the instructions here: https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/allowing-changes-to-a-pull-request-branch-created-from-a-fork

[LeeviLindgren]

@LeeviLindgren Thanks for the PR! Sorry for the delay in reviewing it. I'll make a few review comments now, but can you also edits from maintainers? You can do that by following the instructions here: https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/allowing-changes-to-a-pull-request-branch-created-from-a-fork

Thanks for the comments, I will take a look! Edits from maintainers should be allowed.

[LeeviLindgren]

@avehtari also suggested that it would be nice to support model comparison using these additional metrics. I suggest we try to merge this PR which implements the metrics, and I can start working on the functionality to support comparison. The initial idea I had was to write a method for loo_compare for loo_predictive_error-class (we can also come up with another name). Users could then pass output from the calls to loo_predictive_error to loo_compare. What do you think @jgabry?

[jgabry]

Thanks for the updates. Waiting to hear from @avehtari on the couple of questions in my review comments (about the argument names) but after that we can merge this (assuming the GitHub Actions checks pass).

@avehtari also suggested that it would be nice to support model comparison using these additional metrics. I suggest we try to merge this PR which implements the metrics, and I can start working on the functionality to support comparison. The initial idea I had was to write a method for loo_compare for loo_predictive_error-class (we can also come up with another name). Users could then pass output from the calls to loo_predictive_error to loo_compare. What do you think @jgabry?

Yeah I agree it would be nice to be able to do model comparison with these. My initial thought is that your suggestion of using a class and a loo_compare method seems like a good approach.

[avehtari]

I agree with @jgabry's suggestions

[LeeviLindgren]

• I changed the function to loo_predictive_metric
• Updated the wording in docs to match the updates
• ll argument is now called log_lik
• pred_error argument is now called metric

[jgabry]

Looks good, thanks! Merging now.