imputation/README.md

Imputation track — leaderboard substrate

This directory holds the per-method substrate for the OpenMHC Track-2 (imputation) leaderboard. Each method ships two files:

imputation/<method>.parquet         # per-(user × scenario × channel × subgroup) errors
imputation/<method>.meta.json       # display + diagnostic sidecar

See SCHEMA.md for the exact column / field schema (including the fallback_rate diagnostic).

What's it for

The substrate parquets are the canonical inputs for:

• The OpenMHC HF Space (MyHeartCounts/OpenMHC) — leaderboard_compute.py there downloads these parquets + the sidecars and runs the canonical reducers from imputation_evaluation to produce the live leaderboard table.
• Independent re-aggregation (skill / rank / fairness reducers in src/imputation_evaluation/evaluation/)
• The cluster-bootstrap reference at imputation/bootstrap*/ is reduced from these substrates, so any change here propagates downstream.

Loading

from huggingface_hub import hf_hub_download
import pandas as pd, json

# One method's substrate
parquet = hf_hub_download(
    "MyHeartCounts/OpenMHC-leaderboard-data",
    "imputation/locf.parquet",
    repo_type="dataset",
)
df = pd.read_parquet(parquet)
print(df.shape, df.columns.tolist())

# Display + diagnostic sidecar (incl. fallback_rate)
meta_p = hf_hub_download(
    "MyHeartCounts/OpenMHC-leaderboard-data",
    "imputation/locf.meta.json",
    repo_type="dataset",
)
meta = json.loads(open(meta_p).read())
print(meta)
# -> {"display_name": "LOCF (baseline)", "type": "Statistical", ...,
#     "fallback_rate": 0.0}

Pooled substrate (BCa LOO)

The pooled per-user errors frame across all methods is NOT stored here — it is exactly the concatenation of the per-method parquets:

import glob, pandas as pd
pooled = pd.concat(
    [pd.read_parquet(p) for p in glob.glob("imputation/*.parquet")],
    ignore_index=True,
)
# ~2.5M rows = 148,510 rows/method × 17 methods (or × 16 for the legacy pool)

The bootstrap reference under imputation/bootstrap/ was computed against the 16-method pool (legacy / paper-matching). The sibling imputation/bootstrap_with_dense_weekly/ was computed against the 17-method pool that includes the lsm2_weekly dense variant.

Refreshing

The substrate is produced and uploaded by the OpenMHC code repo:

# (Phase A) Per-method runs land at runs/<method>/{pairs/, per_user_errors.parquet, results.json}
bash jobs/sherlock/imputation_eval/submit_all.sh --no-paper

# (Phase B+C+D) Bootstrap → substrate producer → HF upload (chained)
JID_A=$(sbatch --parsable jobs/sherlock/imputation_eval/run_paper_bootstrap.sbatch)
JID_B=$(sbatch --parsable jobs/sherlock/imputation_eval/run_paper_bootstrap_no_dense.sbatch)
JID_C=$(sbatch --parsable --dependency=afterok:$JID_A \
         scripts/paper_results/imputation/parity/produce_per_method_per_user_errors.sbatch)
JID_D=$(sbatch --parsable --dependency=afterok:$JID_B:$JID_C \
         jobs/sherlock/imputation_eval/upload_leaderboard.sbatch)

The upload step auto-extracts fallback_rate from each method's results.json and threads it into the sidecar without clobbering the existing display fields. See jobs/sherlock/imputation_eval/README.md for the canonical recipe.