Skip to content

Full eeg leaderboards - #30

Merged
TonyBagnall merged 2 commits into
mainfrom
full-eeg-leaderboards
Sep 19, 2026
Merged

TonyBagnall merged 2 commits into
mainfrom
full-eeg-leaderboards

Conversation

@TonyBagnall

Copy link
Copy Markdown
Contributor

Summary

  • What does this PR change?

Checklist

  • I have updated documentation if needed.
  • I have added tests or validation steps if needed.
  • If this PR submits results, it follows results/schema.md.

TonyBagnall and others added 2 commits September 19, 2026 10:50
Full archive. docs/leaderboard_full.md ranks every estimator with a
resample-0 result on all 100 datasets of the Multiverse archive paper's
full-archive comparison: the datasets on which all 17 of its classifiers
completed, now fixed in results/multiverse/paper_datasets.txt. The 17 ranked
are exactly the paper's, with HC2 first at 6.31 and MRHydra second at 6.86 as
in the paper. Below the table, the twelve core-table estimators that have not
finished those 100 are listed with what each is missing, and
pending_full.csv holds the same 508 gaps as estimator,dataset rows to queue
runs from.

EEG. docs/leaderboard_eeg.md does the same over the 26 multivariate datasets of
the EEG archive (eeg2026 less the univariate EpilepticSeizures and Sleep): 14
estimators ranked, 13 pending, 188 gaps in pending_eeg.csv. The averaged
results already in results/eeg, which include the EEG-specific CSP-SVM, R-KNN
and R-MDM, follow as a separate table, marked as not directly comparable.

Both are linked from the README and built by python -m
multiverse.experiments.tables, with sortable HTML versions beside the core one.

To support them:

- Result files now hold every Multiverse dataset an estimator has a resample-0
  result for, not only Multiverse-core: 6041 rows added, none changed. The core
  and UEA tables are unchanged, since they select their datasets.
- CIF-500 and DrCIF-500, the configurations the paper reports as CIF and
  DrCIF, are added. They are withheld from the core table, with the reason
  stated there, so it still reports the default CIF and DrCIF.
- ingest scores a test split that is missing a class. tsml-eval raises on log
  loss and AUROC for those, so the two are computed with the full label set;
  the KERAAL multiclass problems in the 100 need it.
- leaderboard() gains core_sections, to leave the core table's held-out lists
  off other collections' pages, and extra_html.
- A gap outside Multiverse-core with no recorded reason is now reported as not
  run outside Multiverse-core, which is what it is, rather than as reason not
  recorded. This changes the UEA page's missing list, not its table.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A one-line menu under the opening description links the Multiverse-core,
full-archive and EEG leaderboards, replacing the list that sat below the core
table.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
@TonyBagnall
TonyBagnall merged commit d669585 into main Sep 19, 2026
1 of 3 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant