imputation/bootstrap/README.md

Imputation bootstrap reference

This subdir holds the Phase-1 bootstrap reference for the Track-2 (imputation) leaderboard recompute — the long-format per-draw frame that the skill / rank / fairness CIs reduce from.

Layout

imputation/bootstrap/
├── draws.parquet      # per-(method, scenario, channel, subgroup, draw) E/R/rank
└── draws.meta.json    # provenance: seed, n_boot, methods, scenarios, git commit

What's it for

Each row of draws.parquet is one bootstrap draw of one task for one method. Phase-2 aggregators reduce it three ways:

• Skill score S = 1 − exp(mean_r log(R)) — paired against locf; per-(method, scope) mean / SE / percentile-CI across draws.
• Average rank — per-(method, scope) mean of the per-draw cross-method rank, mean / SE / CI across draws.
• Fairness skill score — per-attribute mean-absolute-pairwise-difference disparity ratio D_method / D_baseline, geomean-averaged across tasks, with BCa intervals.

The same draws frame backs all three; only the reducer changes.

Provenance

Generated by scripts/paper_results/imputation/bootstrap_imputation_draws.py in the code repo. The current snapshot was built with:

• seed = 42, n_boot = 1000
• splits = ["test"]
• All 6 imputation scenarios
• 16 methods (see draws.meta.json:methods)
• age_bins = [18, 30, 40, 50, 60], exclude_unknown = false

See draws.meta.json for the exact git commit, method-dirs manifest, and runtime metadata.

Note: no pooled per_user_errors.parquet

The BCa LOO substrate (per-user errors pooled across all methods) is not stored here — it is exactly the concatenation of the per-method substrate files one level up:

import glob, pandas as pd
pooled = pd.concat(
    [pd.read_parquet(p) for p in glob.glob("imputation/*.parquet")],
    ignore_index=True,
)
# 2,376,160 rows = 148,510 rows/method × 16 methods

The same provenance (seed, n_boot, method list, git commit) lives in draws.meta.json here.

Loading

from huggingface_hub import hf_hub_download
import pandas as pd, json

draws_path = hf_hub_download(
    "MyHeartCounts/OpenMHC-leaderboard-data",
    "imputation/bootstrap/draws.parquet",
    repo_type="dataset",
)
meta_path = hf_hub_download(
    "MyHeartCounts/OpenMHC-leaderboard-data",
    "imputation/bootstrap/draws.meta.json",
    repo_type="dataset",
)
draws = pd.read_parquet(draws_path)
meta = json.loads(open(meta_path).read())
print(meta["seed"], meta["n_boot"], len(meta["methods"]))

See SCHEMA.md for the full column spec.

Sibling: 17-method variant

imputation/bootstrap_with_dense_weekly/ holds the same reducer applied to a 17-method pool that adds lsm2_weekly (dense 7-day LSM-2). Skill / fairness numbers per method are identical to this canonical variant (they're pairwise vs locf); only average-rank values shift because the comparison pool grew. See that dir's README for when to use which.

Uploaded with

tools/upload_leaderboard_bootstrap.py in the code repo.