PhaaNe/checkpoint_KKG

KGGuard+W: checkpoints and reproduction artifacts

This repository holds every artifact needed to reproduce the results of KGGuard+W, a method that detects erroneous facts in knowledge graphs. It covers DBpedia Accuracy and FactBench. Code: GitHub coderkhongodo/KGG (see its README).

KGGuard+W does not train a neural network of its own. Seven label-free evidence sources are calibrated into log-likelihood ratios around a hierarchical predicate prior, then combined by a cross-fitted, monotone boosted GAM:

• C: SHACL constraints
• S: structural anomaly (Isolation Forest)
• S_sup: relation-conditional structural support
• M: sentence similarity
• T: text mention
• N: NLI on entity descriptions
• W: NLI on retrieved web passages

The learned state, i.e. the "checkpoint", is the fitted fusion model. Missing evidence contributes exactly 0, and per-source contributions are exact.

Contents

Path	What
checkpoints/<dataset>/<run>/kgguard_plus_w_seed<s>.joblib	final fusion model (KGGuard+W) per dataset × run × seed; .json = readable description
checkpoints/<dataset>/<run>/kgguard_plus_kg_only_seed<s>.joblib	the same model without web evidence (KGGuard+)
checkpoints/verified_metrics.csv	test metrics of every checkpoint; they equal the reported ones (refit check ≤ 1e-9)
outputs/<dataset>/<run>/scores/	every cached evidence and baseline score, including LLM zero-shot, FactCheck (DKA, GIV-F, RAG, consensus), TransE, ComplEx, CAGED, PaTyBRED, SDValidate
data/raw, data/interim, data/interim_ms	raw and processed facts, frozen splits, context graphs; interim_ms = the 5 fresh splits
data/cache	SPARQL / HTTP responses, encoder embeddings, NLI and LLM / FactCheck answers, retrieved web passages
reference/<dataset>/<run>/	reported results (E1–E7) used for comparison; reference/summaries/: multi-split, robustness, label efficiency, paper tables
explanations/<dataset>/<run>/	alert lists with per-source explanations, test predictions
MANIFEST.json	sha256 of every file, library versions, code commit

Runs:

• v7 is the design split, with seeds 13, 42 and 2027.
• ms7_s101 … ms7_s505 are five fresh grouped splits, each with seed 13.

Reproduce

git clone https://github.com/coderkhongodo/KGG.git && cd KGG
pip install -e ".[dev,hf]"
PYTHONPATH=src python scripts/reproduce_from_hf.py --local-dir hf_checkpoint --full-e1 --ablation

The script:

1. downloads this repository and verifies every file against MANIFEST.json;
2. reloads every checkpoint, scores the test facts and compares AP, AUROC, F1 and BAcc with the reported values;
3. refits every model from the downloaded scores and compares again;
4. with --full-e1, re-runs the full main comparison (all baselines, 3 seeds, bootstrap);
5. with --ablation, re-runs the ablation of the proposed method (15 variants, design split and 5 fresh splits).

No network access, LLM or GPU is needed.

Verified on a fresh clone with an anonymous download, on CPU (32 threads):

Check	Result
sha256 of all 3,860 files	0 mismatches
32 checkpoints, inference	AP / AUROC / F1 / BAcc identical (0 difference)
32 checkpoints, retrained from the downloaded scores	identical
full E1 (29 methods x 3 seeds x 2 datasets)	identical
ablation (240 rows)	identical

Checkpoints are pickles (joblib). Load them only if you trust the source, preferably with the same scikit-learn version:

• Python 3.12
• scikit-learn 1.9.1
• numpy 2.5.4
• pandas 3.1.1

Main results (test AP, mean of 3 seeds)

	DBpedia Accuracy	FactBench
best published baseline	0.257 (LLM zero-shot)	0.834 (FactCheck-RAG)
CAGED (CIKM'22)	0.142	0.533
KGGuard+W	0.470	0.928

Data sources and licences

Redistributed for reproducibility, under the licences of the original sources:

• DBpedia Accuracy (Marchesin, Silvello, Alonso, HCOMP 2024) and DBpedia context data: CC BY-SA.
• FactBench (DeFacto, AKSW).
• Web passages retrieved from the FactCheck corpus (SIGIR'25 / EDBT'26).

Please cite the original datasets when you use them.