MoshinAli/aefs-cpdp

AEFS — Adaptive Explainable Feature Selection for Cross-Project Defect Prediction

AEFS is a four-stage feature-selection framework for Cross-Project Software Defect Prediction (CPDP): rapid uninformative pruning, sequential forward selection, SHAP-based refinement and cross-project stability validation. It is evaluated on four benchmark pools (AEEEM, TeraPromise, NASA MDP, ReLink — 80 project-versions) with Leave-One-Project-Out (LOPO) cross-validation.

This repository is the reference implementation and the complete experimental record: source code, benchmark data, raw result tables and analysis scripts.

Headline results

• Feature count reduced by 75–89% on all four pools while staying competitive with the SOTA baseline BorutaSHAP (which retains 100% of the features): |ΔMCC| ≤ 0.077 under identical LOPO protocols.
• ReLink: MCC 0.228 with 78.2% reduction. NASA MDP: all methods are near-zero MCC (defect rates of only 11–13%).
• Statistics are reported to Q1 standards: Wilcoxon signed-rank with Cliff's δ effect sizes and Holm–Bonferroni correction within each pool, Friedman + Nemenyi post-hoc, and explicit power discussion (3 of 4 pools have < 6 projects).
• Fold runtime 66.73 s on AEEEM (65× AllFeatures, 31× the next method) — practical for offline pre-deployment analysis.
• Seed sensitivity quantified over 5 seeds: per-target MCC std 0.033 (AEEEM) / 0.053 (TeraPromise), against an across-target std of 0.072 / 0.135 — the reported differences are not seed artefacts.

Repository layout

AEFS_Pipeline.ipynb               main pipeline (stages 1-4, LOPO, LightGBM/RF/SVM)
AEFS_Extended_Eval.ipynb          extended evaluation (5 methods x 3 classifiers)
AEFS_ReLink_NASA_Expansion.ipynb  ReLink + NASA MDP pools
AEFS_Sensitivity_Analysis.ipynb   SHAP / stability threshold sweeps, 5-seed runs
scripts/                          pipeline + analysis + replication scripts (see below)
data/                             aeeem/ nasa/ relink/ tera/ CSVs (+ AEFS_data.zip mirror)
results/experiments/              raw per-fold rows for every experiment (sharded + merged)
results/analysis/                 descriptive stats, RQ1-RQ6 CSVs (rq1_rq2..rq6)
results/*_stage4_results.csv      per-pool x per-method summaries (LightGBM)
results/*_extended_results.csv    per-pool x per-method summaries (RF / SVM)
results/registry.json             machine-readable experiment states
sensitivity/results/              authoritative threshold sweeps, runtime, convergence
sensitivity/results_kaggle/       independent replication (unpinned, 16 workers)
sensitivity/results_pinned/       independent replication (pinned library versions)
sensitivity/results_multiseed/    5-seed runs
figures/                          matplotlib exports written by the notebooks
requirements.txt                  pinned dependencies

The four benchmark pools are the standard public CPDP datasets: AEEEM, TeraPromise, NASA MDP and ReLink. data/ holds them as per-project CSVs in a uniform schema; AEFS_data.zip is an identical archive of the same files.

Manuscript sources and compiled PDFs are deliberately not part of this repository; everything needed to re-run the experiments and independently re-derive every reported number is here.

Reproducing the results

pip install -r requirements.txt
jupyter lab                    # run the four notebooks in order, or

Everything below is scripted and idempotent:

python scripts/regen_sensitivity.py --pools aeeem tera   # threshold sweeps, runtime, convergence
python scripts/compute_stats.py                          # per-pool tables + Wilcoxon/Cliff's delta
python scripts/stats_q1.py                               # Holm-Bonferroni, Nemenyi, power
python scripts/analyses.py                               # RQ1-RQ5 statistics from the raw rows
python scripts/analyze_extended.py                       # extended 5-method x 3-classifier analysis
python scripts/cmp_replication.py                        # replication vs. authoritative numbers
python scripts/thread_test.py                            # thread-count nondeterminism check

results/analysis/rq6.csv is the machine-checkable link between the raw experiment rows and the narrative claims: every claim maps onto verified aggregate values, so a reader can re-derive the paper's numbers directly from the CSVs in this repository.

requirements.txt pins the exact versions used for every reported number in results/ and sensitivity/results/.

Independent replication

The pipeline was re-run on a second machine (Linux, 4 CPUs, Python 3.13.15, scikit-learn 1.6.1, lightgbm 4.6.0) both with the natural library versions and with the authoritative ones pinned. Highlights: feature-count selections are identical on 268/272 threshold-sweep cells, and the seed-variance statistics reproduce (0.033 / 0.054 per-target std). The three replication runs are in sensitivity/results_kaggle/, sensitivity/results_pinned/ and sensitivity/results_multiseed/; scripts/cmp_replication.py compares them against the authoritative tables.

A mirrored copy of the code, data and results is available as the Hugging Face dataset MoshinAli/aefs-cpdp.

Citing this repository

Cite via CITATION.cff (the Cite this repository button on GitHub). Release history is in RELEASE_NOTES.md; the DOI above is the Zenodo concept DOI, which always resolves to the latest release. Release v1.0.0 is the frozen pre-submission snapshot (10.5281/zenodo.23274379); release v1.1.0 carries the final author metadata and figure fixes.

License

MIT — see LICENSE.