minhy112/calibkgc-icta2026

CalibKGC

Artifacts for CalibKGC: Confidence Calibration and Selective Prediction for Knowledge Graph Completion on WN18RR and FB15k-237.

The repository contains the code, reciprocal-augmented datasets, PyKEEN training outputs, exported ranking scores, calibration/selective prediction results, figures, and the compiled Springer LLNCS paper PDF.

Paper

• LaTeX source: paper/main.tex
• References: paper/references.bib
• Compiled PDF: paper/main.pdf

Main reported PyKEEN runs use TransE and RotatE, 200 epochs, embedding dimension 200, batch size 2048, 32 negatives, Adam learning rate 5e-4, and seed 7.

Data

Expected dataset layout:

data/
  wn18rr/
    entity2id.txt
    relation2id.txt
    train.txt
    valid.txt
    test.txt
  fb15k-237/
    entity2id.txt
    relation2id.txt
    train.txt
    valid.txt
    test.txt

Each triple file uses integer IDs:

head relation tail

Quick Commands

Collect PyKEEN ranking metrics:

python scripts/collect_pykeen_metrics.py \
  --root results/pykeen_runs \
  --output results/tables/pykeen_ranking.csv

Train one PyKEEN model:

python scripts/train_pykeen.py \
  --dataset wn18rr \
  --model RotatE \
  --seed 7 \
  --epochs 200 \
  --dim 200 \
  --batch-size 2048 \
  --num-negatives 32 \
  --lr 0.0005 \
  --output-dir results/pykeen_runs

Export filtered tail-ranking score files:

python scripts/export_pykeen_scores.py \
  --dataset wn18rr \
  --model-path results/pykeen_runs/wn18rr/rotate/seed_7/trained_model.pkl \
  --split test \
  --batch-size 64 \
  --top-k 50 \
  --output results/pykeen_scores/wn18rr/rotate/seed_7/test_scores.csv

Fit calibration methods and evaluate selective prediction:

python scripts/calibrate.py \
  --scores-dir results/pykeen_scores/wn18rr/rotate/seed_7 \
  --dataset wn18rr \
  --model rotate \
  --seed 7 \
  --output-dir results/pykeen_calibration

Aggregate all completed runs:

python scripts/aggregate_results.py \
  --raw-root results/pykeen_scores \
  --calib-root results/pykeen_calibration \
  --output-dir results/tables

Compile the paper:

cd paper
latexmk -pdf -interaction=nonstopmode main.tex

Final Reported Results

Ranking metrics are in results/tables/pykeen_ranking.csv.

Calibration and selective prediction summaries are in:

• results/tables/calibration_summary.csv
• results/tables/selective_summary.csv
• results/tables/relation_miscalibration_wn18rr_rotate_groups.csv
• results/tables/relation_miscalibration_fb15k-237_rotate_groups.csv

Figures used by the paper are in results/figures/.

Notes

The results/probe*, results/smoke*, and results/raw_scores directories contain earlier exploratory runs and are preserved for reproducibility, but the paper tables use the PyKEEN outputs under results/pykeen_runs, results/pykeen_scores, and results/pykeen_calibration.

To reproduce the main matrix:

for dataset in wn18rr fb15k-237; do
  for model in TransE RotatE; do
    python scripts/train_pykeen.py \
      --dataset "$dataset" \
      --model "$model" \
      --seed 7 \
      --epochs 200 \
      --dim 200 \
      --batch-size 2048 \
      --num-negatives 32 \
      --lr 0.0005 \
      --output-dir results/pykeen_runs
  done
done