Link Prediction or Perdition: the Seeds of Instability in Knowledge Graph Embeddings
Guillaume Méroué, Fabien Gandon, Pierre Monnin · 2026
Read Link Prediction or Perdition: the Seeds of Instability in Knowledge Graph Embeddings by Guillaume Méroué, Fabien Gandon, Pierre Monnin (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-03365.txt.
Description
Méroué, Gandon, and Monnin test whether knowledge graph embedding models give stable answers. Such models are the main tool for link prediction, which fills gaps in knowledge graphs. Standard grading favors rank scores such as MRR. It skips how random seeds shake the results. Those scores can also hide wobble in single predictions and in the shape of the embedding space.
The team runs a systematic stability study across many models and datasets. High scorers in fact diverge at the level of single triples. Their embedding spaces vary a lot from run to run. The team isolates the random factors: initialization, triple order, negative sampling, dropout, and hardware. Each one alone induces instability of like size. Better MRR does not promise better stability for a fixed model. Voting, a known aid, helps only a bit. The authors warn that current grading hides a reliability problem for graph completion.
Catalog record
| CREATOR | Guillaume Méroué, Fabien Gandon, Pierre Monnin |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 43 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2606.03365