Knowledge Graph Entity Alignment (Master's Internship)

Institution: LISN – Laboratory for Interdisciplinary Numerical Sciences, University of Paris-Saclay
Degree: Master’s Internship
Dates: June–July 2024
Author: Nassim Arifette
Supervisors: Dr. Fatiha Saïs, Olivier Inizan

Project OverviewCopy link to section

During a 2-month internship at LISN (Univ. Paris-Saclay), I conducted an in-depth, head-to-head comparison of knowledge graph embedding (KGE) models for entity alignment on a real-world, tri-source dataset integrating YAGO, Wikidata, and DBpedia. I evaluated representative translational (TransE, TransH), bilinear (DistMult, ComplEx), and convolutional (ConvE) approaches within a unified experimental framework.

Research Context & MotivationCopy link to section

Multiple public KGs overlap but diverge in schema, language, and graph structure, creating the need to identify entries that refer to the same real-world entity (entity alignment). This work addresses that need by comparing KGE families not only on ranking accuracy but also on semantic preservation—how well embeddings retain graph-structural meaning crucial for reliable alignment.

Data & Experimental SettingCopy link to section

• Integrated KG (entertainment domain): 47,303 RDF triples, 5,807 entities, 16 relations (≈0.14% density; highly sparse).
• Splits: 80% train / 10% validation / 10% test; filtered evaluation protocol for link prediction.
• Hardware/Framework: PyTorch 1.13, CUDA 11.7, NVIDIA V100, 5 seeds for robustness.

Models EvaluatedCopy link to section

• Translational: TransE, TransH
• Bilinear: DistMult, ComplEx
• Convolutional: ConvE (2D CNN over entity–relation reshaped embeddings)

Key Quantitative FindingsCopy link to section

Link Prediction (mean ± std across 5 runs)

Highlights

• ConvE achieves 22% relative MRR improvement over TransE (0.381 vs 0.312).
• Clear accuracy–efficiency trade-off: ConvE is most accurate; TransE is the most computationally efficient.

Computational Efficiency (indicative)

• Time/epoch (s): TransE 12.3 → ConvE 38.6; peak GPU ≈ 1.2GB (TransE) to 4.7GB (ConvE).
• “Time per 0.01 MRR” quantifies cost–benefit; ConvE costs ~20.26s per +0.01 MRR vs TransE ~7.88s.

Entity Alignment & Semantic PreservationCopy link to section

Beyond ranking, I evaluated alignment quality (cosine similarity NN retrieval) and semantic fidelity (correlation between 1-hop Jaccard similarity and embedding cosine similarity).

• Alignment: ConvE achieves the best P@1/R@1/F1@1 and top-K performance among evaluated models.
• Semantic Preservation: ConvE exhibits the strongest correlation (r ≈ 0.72) between graph-structural similarity and embedding similarity (vs DistMult ≈ 0.41), showing that high link-prediction accuracy does not always imply semantically faithful embeddings.

Practical RecommendationsCopy link to section

• For web-scale efficiency: Start with TransE; consider TransH if many complex relations justify extra cost.
• For maximum precision: Choose ConvE when higher computational budget is acceptable.
• Avoid DistMult on asymmetric relations (e.g., directedBy, parentOf).

Methodology SnapshotCopy link to section

• Unified training/evaluation framework with negative sampling, consistent filtered metrics (MR, MRR, Hits@K), and early stopping on validation MRR.
• Relation-type analysis shows ConvE’s advantage on many-to-many (N–N) relations (e.g., acted_in).

Research ImpactCopy link to section

• Scientific: Demonstrates a clear expressiveness hierarchy (DistMult < TransE < TransH < ComplEx ≪ ConvE) and evidences the semantic preservation paradox—ranking gains don’t always translate to faithful semantics.
• Practical: Offers guidelines for choosing a model under accuracy and efficiency constraints. The effect on downstream manual curation was not measured in this study.

Technical ImplementationCopy link to section

• Stack: PyTorch 1.13, CUDA 11.7; Intel Xeon Gold 6248R; NVIDIA V100; 256GB RAM; 5 random seeds for robustness.
• Reproducibility: Documented preprocessing, splits, and seed settings; modular code layout for data, models, and evaluation.

Illustrative Project Structure

entity_alignment/
├── data/ # loader.py, preprocessor.py, sampler.py
├── models/ # transe.py, transh.py, distmult.py, complex.py, conve.py
├── evaluation/ # metrics.py, evaluator.py
└── main.py

Personal Learning OutcomesCopy link to section

• Technical: Deep learning for graph data; fair benchmarking; significance-aware evaluation.
• Research: Designing controlled experiments; connecting empirical results to theory; semantic fidelity analysis beyond standard metrics.

Future DirectionsCopy link to section

• Evaluate on web-scale KGs; include RotatE and GNN-based methods;
• Develop semantic-aware objectives to explicitly preserve graph structure;
• Extend cross-lingual alignment and robustness analyses.