missing_prediction / README.md
phamhcuong's picture
Complete Experiment 2 at rho=0.3: Scenario 1, Scenario 2, A/B F1 and token accounting
b247547 verified
|
Raw
History Blame Contribute Delete
5.28 kB
---
license: other
tags:
- missing-data
- imputation
- benchmark
- healthcare
---
# Missing-Data Prediction Benchmark — results
Two experiments on medical datasets under simulated missingness (MCAR/MAR/MNAR,
rates 0.1–0.5, seeds 42/1/7), reported as **mean [95% CI]**.
This release contains the complete Information-Acquisition benchmark at a uniform missing
rate of 0.3. Scenario 2 includes local AFA-engine results, curves, seed-aggregated CI
tables, qualitative examples, classification AUROC, and per-engine cached/uncached token
accounting across five classification datasets (`arrhythmia`, `breast_cancer`, `gallstone`,
`heart_disease`, `parkinsons`) and four regression datasets (`diabetes`, `nhanes`,
`who_life_expectancy`, `medical_cost`).
Scenario 1 uses `gpt-5.6-terra` to select features and predict directly. It is complete for
the same classification and regression dataset sets. Classification Scenario 1 reports
accuracy, F1, precision, and recall; AUROC and P(true) are intentionally unavailable because
this model does not expose token logprobs. Regression Scenario 1 predicts a continuous numeric
target and reports RMSE, MAE, and R2. The classification Scenario 1-vs-Scenario 2 A/B report
uses accuracy, F1, and exact matched-sample token accounting; it does not claim an LLM AUROC.
MIMIC-III is not part of Experiment 2 because the tabular AFA runner does not support the
derived MIMIC benchmark. Experiment 1 contains only aggregate metrics for derived tabular
MIMIC length-of-stay prediction; no patient-level records, raw MIMIC data, or trained models
are published.
### Classification Scenario 1 vs Scenario 2
| dataset | mech | LLM acc | AFA acc | LLM F1 | AFA F1 | best AFA | LLM tokens | S2 cached | S2 uncached | ↓ cached | ↓ uncached |
|---|---|--:|--:|--:|--:|---|--:|--:|--:|--:|--:|
| arrhythmia | MAR | 0.469 | 0.736 | 0.411 | 0.734 | random_forest | 4,638,941 | 31,992 | 417,730 | **99.3%** | 91.0% |
| arrhythmia | MCAR | 0.44 | 0.74 | 0.408 | 0.738 | random_forest | 5,432,001 | 31,992 | 404,444 | **99.4%** | 92.6% |
| arrhythmia | MNAR | 0.476 | 0.707 | 0.392 | 0.704 | random_forest | 2,869,316 | 31,992 | 414,435 | **98.9%** | 85.6% |
| breast_cancer | MAR | 0.921 | 0.977 | 0.921 | 0.977 | smart | 348,937 | 4,835 | 56,524 | **98.6%** | 83.8% |
| breast_cancer | MCAR | 0.927 | 0.98 | 0.927 | 0.979 | mim_smim | 373,315 | 4,191 | 13,165 | **98.9%** | 96.5% |
| breast_cancer | MNAR | 0.927 | 0.974 | 0.927 | 0.974 | smart | 348,073 | 4,835 | 67,905 | **98.6%** | 80.5% |
| gallstone | MAR | 0.495 | 0.771 | 0.483 | 0.769 | hist_gb | 823,103 | 5,990 | 50,355 | **99.3%** | 93.9% |
| gallstone | MCAR | 0.495 | 0.755 | 0.482 | 0.755 | mim_smim | 830,642 | 6,043 | 43,630 | **99.3%** | 94.7% |
| gallstone | MNAR | 0.51 | 0.771 | 0.499 | 0.77 | random_forest | 814,237 | 6,043 | 48,555 | **99.3%** | 94.0% |
| heart_disease | MAR | 0.7 | 0.844 | 0.698 | 0.844 | smart | 197,635 | 2,068 | 18,610 | **99.0%** | 90.6% |
| heart_disease | MCAR | 0.694 | 0.856 | 0.693 | 0.855 | mim_smim | 208,505 | 2,068 | 11,775 | **99.0%** | 94.4% |
| heart_disease | MNAR | 0.694 | 0.833 | 0.693 | 0.833 | smart | 198,665 | 2,068 | 18,187 | **99.0%** | 90.8% |
| parkinsons | MAR | 0.761 | 0.897 | 0.776 | 0.889 | hist_gb | 159,996 | 3,368 | 19,475 | **97.9%** | 87.8% |
| parkinsons | MCAR | 0.752 | 0.889 | 0.767 | 0.887 | hist_gb | 155,967 | 3,322 | 15,660 | **97.9%** | 90.0% |
| parkinsons | MNAR | 0.744 | 0.872 | 0.759 | 0.873 | xgboost | 164,584 | 3,373 | 14,908 | **98.0%** | 90.9% |
## How to read (no download needed)
Every CSV opens in HuggingFace's Dataset Viewer (sortable, in-browser):
- `experiment1_prediction/prediction_classification_long.csv` / `prediction_regression_long.csv`**one tidy row per (method, dataset, mechanism, rate, seed)** with all metrics + status. This is the clean, flat view of every run (no folder-diving).
- `experiment1_prediction/classification_mean_ci.csv` / `regression_mean_ci.csv` — the same, collapsed to mean [95% CI] over seeds (blank cells carry a `Status`).
- `experiment2_acquisition/acquisition_ab_report.csv` — matched-sample classification Scenario 1 vs Scenario 2 accuracy and token comparison.
- `experiment2_acquisition/curves_long.csv` — per-budget Scenario 2 trajectories and Scenario 1 trajectories for both tasks.
- `experiment2_acquisition/acquisition_curves_ci.csv` / `acquisition_final_ci.csv` — seed-aggregated metrics for both scenarios.
- `experiment2_acquisition/scenario2_method_token_accounting.csv` — per-engine cached and uncached Scenario-2 token accounting.
- `DATA_DICTIONARY.md` — column definitions and status meanings.
- `raw_prediction/` — per-cell raw Exp-1 metric JSON (in-scope datasets, canonical layout, see raw_prediction/README_LATEST.md).
- `raw_acquisition/` — per-cell raw complete Scenario-1 LLM and Scenario-2 AFA outputs (no model weights).
## Data-use / MIMIC-III notice
One regression dataset (`mimic_lengthofstay`) is derived from **MIMIC-III**, credentialed
under the PhysioNet Data Use Agreement. This repository contains only **aggregate derived
metrics** — no patient-level records, no raw MIMIC data, and no models trained on MIMIC.
Generated 2026-09-04 by `scripts/reporting/prepare_hf_upload.py`.