Reproducibility Artefacts and Experimental Results for Post-hoc Worst-Group Robust Ensemble Weighting
Description
This dataset contains the reproducibility artefacts and derived experimental results associated with the study “Post-hoc Worst-Group Robust Ensemble Weighting: A Reproducible Empirical Evaluation on Tabular Classification Benchmarks”. The deposit includes the code, configuration files, experimental outputs, statistical analysis files, and documentation needed to reproduce the reported results. The public datasets used in the study are not redistributed here and should be obtained from their original sources.
Files
Steps to reproduce
1. Obtain the public datasets from their original sources. Raw datasets are not included in this deposit. Place the downloaded data under data/raw/<dataset_name>/ according to supplementary/reproducibility.md. The synthetic dataset is generated by scripts/generate_synthetic_dataset.py. 2. Install the project dependencies using the supplied environment/configuration files. 3. To reproduce the reported tables and figures from the committed campaign checkpoints, run from the archive root: python scripts/run_stage5_campaign.py --aggregate-only This regenerates the tables and figures under results/campaign/tables/ and results/campaign/figures/ using the documented statistical pipeline. 4. To run the full campaign from scratch with the documented trial count and analysis seed: python scripts/run_stage5_campaign.py --trials 30 --analysis-seed 20260726 The campaign uses R = 30 trials per dataset. German Credit retains 30 checkpoints, of which 5 pass the N1 validity check and are counted; 25 are retained as excluded checkpoints. Folktables ACS Income is a documented data-pipeline exclusion. 5. For a single-dataset Stage 1 sanity check, run: python scripts/run_stage1_pipeline.py --base-config configs/base_config.yaml --dataset-config configs/dataset/adult.yaml 6. Detailed preprocessing, configuration, trial accounting, statistical procedures, and reproducibility information are provided in the included supplementary documentation.