Sparse Hypergraph Interaction Learning

Date/time: 2026-07-30 19:53 +03:00 Tool: Codex Model, if known: GPT-5.6 Extra High (xhigh, user-attested) Operation ID: shil-ijmlc-public-repo-sync-20260730-1953

Replication package for the diagnostic sparse-interaction support-selection manuscript.

Status: public replication package. The repository includes the initial experiments and the final stability-calibration, Dry Bean, and fixed-C artifacts used by the manuscript. It also provides pinned requirements, source and output hashes, per-seed completion records, citation metadata, licensing notices, and non-compute integrity checks.

Two bundles live here

Date/time: 2026-09-22 01:40 +03:00 Tool: Cowork-Claude Model, if known: claude-opus-5 (effort max) Operation ID: shil-informatica-retarget-20260921

• replication_package/ — the earlier implementation, described by an earlier submission. It was not modified when the second bundle was added; every one of its files is byte-identical to what this repository already published.
• ranker_policy_benchmark/ — the experiment bundles behind the current manuscript, A Transparent Ranker–Policy Benchmark for Pair and Triple Interaction-Support Selection: the original frozen benchmark, the ranker-family follow-up, the controlled extension, and the cross-environment check, with a SHA-256 for every file. Its own README.md states what those records do and do not establish, and records the two files that are stored gzipped here because of GitHub's single-file size limit.

Author

Özkan Canay is the sole and corresponding author. He is affiliated with the Department of Information Systems and Technologies, Faculty of Computer and Information Sciences, Sakarya University, Esentepe Campus, 54050 Serdivan, Sakarya, Türkiye. Email: canay@sakarya.edu.tr. ORCID: 0000-0001-7539-6001.

Claim Boundary

This package supports the narrowed diagnostic claim: SHIL is evaluated as a fixed-budget support-selection and reporting procedure for sparse pair/triple interaction regimes. It does not claim broad tabular state-of-the-art performance, general method superiority over direct sparse polynomial logistic regression, or superiority over budget-matched L1 top-k coefficient ranking.

Contents

• code/: reproduction, follow-up, scale-stress, Q1 numeric-update, and bounded-sensitivity scripts.
• replication_package/results/: frozen and Q1 follow-up result summaries.
• replication_package/calibrated_runs/: the hash-locked stability-calibration protocol, twelve-scenario synthetic artifacts, Dry Bean artifacts, fixed-C sensitivity artifacts, source/configuration files, and per-seed records.
• ranker_policy_benchmark/: the current manuscript's bundles — original frozen benchmark, ranker-family follow-up, controlled extension, cross-environment check, figure sources, and SOURCE_MAP.json hashes. See its own README.md.
• docs/: claim boundary and release boundary.
• data_sample/: no raw benchmark data are bundled; public datasets are reconstructed through loaders.
• requirements.txt: pinned environment from the verified local run.
• run_smoke.ps1: dependency, file-presence, and package-local evidence-schema check; it does not run an experiment.
• run_all.ps1: review-oriented wrapper around the same non-compute integrity checks.
• LICENSE and LICENSE-CONTENT.md: software and content license boundaries.
• CITATION.cff: repository citation metadata.

Quick Start

python -m venv .venv
.\.venv\Scripts\Activate.ps1
python -m pip install --upgrade pip
python -m pip install -r requirements.txt
.\run_smoke.ps1

The smoke test verifies dependencies, required files, SHA-256 entries, calibrated run row counts, per-seed completion records, and selected evidence schemas. It does not regenerate evidence. The full calibrated experiments are intentionally separated from this check because they require long CPU runs.

How to rerun

Use a fresh output directory for each run. The commands below are documented from the packaged entry points; they were not executed for this README update. run_smoke.ps1 and run_all.ps1 run integrity checks only.

Current manuscript: ranker-policy benchmark

Follow the bundle's rerun instructions for the v5 benchmark, v6 follow-up, corrected mta006 extension, and vps005 cross-environment check. That section supplies commands, configuration files, seeds, recorded environments, and recorded runtimes for each bundle.

The public snapshot has incomplete execution packaging for v5, mta006, and vps005: required freeze/preflight artifacts are absent, and the extension producers expect a different directory layout. Their commands are conditional on restoring the documented execution prerequisites. The supplied records remain available for inspection; a full rerun from a clean clone is not yet verified. The excluded mta005 attempt is provenance only.

Earlier calibrated package

The earlier package uses the configuration and source under replication_package/calibrated_runs/. Install each run's requirements_locked.txt in a separate environment. The repository-root requirements.txt also includes dependencies for the older auxiliary scripts and the PowerShell smoke check.

These entry commands also retain original relative-input requirements. SC-001 resolves its protocol at replication_package/MD/02_design/method_extension_protocol_20260717.md, whereas the bundled protocol is under the run's protocol/ directory. The fixed-C runner requires its protocol under inputs/ and the original, hash-matching UCI ZIP at cache/uci_602/dry_bean_dataset.zip before execution. Its bundled protocol is under protocol/; the raw ZIP is not redistributed. Restore those expected paths from the matching protocol and dataset sources before using the entry commands; keep the source/configuration hash checks.

From replication_package/calibrated_runs/EXP-SHIL-SC-001/:

python -m pip install -r requirements_locked.txt
python src/sc_shil_experiment.py --config config/full_config.json --mode full-synthetic --output reproduced/full-synthetic
python src/analyze_results.py --input reproduced/full-synthetic --output reproduced/full-synthetic-analysis
python src/sc_shil_experiment.py --config config/full_config.json --mode full-real --output reproduced/full-real

From replication_package/calibrated_runs/EXP-SHIL-FIXED-C-001/:

python -m pip install -r requirements_locked.txt
python src/fixed_c_experiment.py --config config/full_config.json --mode full-real --output reproduced/full-real

The synthetic seeds in the SC-001 configuration are 1103, 1129, 1151, 1171, 1181, 1201, 1213, 1231, 1249, and 1277. Both Dry Bean configurations use 11, 23, 37, 53, 71, 89, 101, 131, 157, and 181. Dry Bean requires network access to the hash-checked UCI archive. The saved SC-001 synthetic environment.json records Windows 11, Python 3.12.12, and scikit-learn 1.8.0; the fixed-C seed-11 environment records Linux/aarch64, Python 3.12.3, and scikit-learn 1.8.0. These are recorded execution environments, not claims of testing on other systems. Runtime for a fresh execution of these commands: not measured. See the per-run READMEs, configurations, and environment records for the historical execution context.

Data Reconstruction

Synthetic datasets are generated by the scripts. Breast cancer and wine are loaded through scikit-learn. Diabetes OpenML 37 is fetched through OpenML by data id. Dry Bean is retrieved from UCI dataset 602 and checked against the recorded archive hash. Raw public benchmark data are not bundled in this repository.

Licensing

• Source code, execution scripts, and software-oriented files are available under the MIT License in LICENSE.
• Documentation and generated result, manifest, selected-edge, figure, and synthetic-example artifacts are available under CC BY 4.0 as specified in LICENSE-CONTENT.md.
• Third-party dataset and dependency rights remain with their respective providers. Raw public benchmark datasets are not redistributed.

Citation

Use the repository metadata in CITATION.cff and identify the commit or release used for an analysis.