InteractiveVLDB 2025 Workshops · LS-NSL
Poderoso: embeddings and expert rules, completing knowledge graphs together
Abelardo Carlos Martínez Lorenzo, Alexander Perfilyev, Volker Markl, Martha Clokie, Thomas Sicheritz-Ponten, Zoi Kaoudi · TU Berlin, University of Copenhagen, University of Leicester, IT University of Copenhagen
Knowledge graphs are always incomplete. Embedding models guess missing links from patterns in the data, but fail where data is sparse; experts know rules the data doesn't show. Poderoso lets the two work in a loop, with any embedding model and any reasoner.
- MRR for HolE on LUBM (0.09 → 0.27)
- 3×MRR for HolE on LUBM (0.09 → 0.27)
- better MRR than pLogicNet, and 1.5× better than KALE
- 3.5×better MRR than pLogicNet, and 1.5× better than KALE
- KGE model or reasoner can be plugged in
- AnyKGE model or reasoner can be plugged in
Where data-driven link prediction runs out
Knowledge graph embeddings place entities and relations in a vector space and score whether a missing triple is likely. Rule mining learns logical patterns from the graph. Both only know what the graph already contains, so they work in dense regions and struggle in sparse ones.
Imagine Bob, a bioinformatician working on bacteriophage therapy. He knows facts from the field that his datasets don't contain, for instance that a certain type of phage can burst the bacterial cell. No embedding will learn that from data that doesn't show it.
Earlier hybrids (KALE, pLogicNet) weave rules into one specific embedding model, which makes them hard to adapt and unable to use ontologies or expert rules.
Neural + symbolic
Embeddings
generalise from patterns; weak where the graph is sparse
Rules
encode what experts know; can only deduce from what is there
Loosely coupled, each one feeds the other: embeddings propose plausible links, rules deduce consequences, and the new facts retrain the embeddings.
Two engines, one loop
The neural engine trains a KGE model on the graph. Entities that share relations end up close; a relation is a direction in the space.
facts in the graph
13 → 13
accepted from KGE
–
derived by rules
0
therapy candidates
–
Graph and rules R1–R5 from Figure 3 and Section 3.2; the reasoning is computed live. Embedding positions, sampled candidates and their scores are illustrative: the paper does not list the KGE outputs. Compare “KGE + rules” with “Rules only”: without the two predicted links, no therapy candidate can be derived.
The hard part is asking the embedding model the right questions. A model can only score triples you give it, and the space of possible triples is enormous:
DBPedia20k
20,143 entities · 12 relations · 120,000 triples
4.9 billion
possible triples (N² × R) to score if nothing were sampled
The graph itself holds 2.5e-3% of them.
LUBM-2
53,860 entities · 32 relations · 268,136 triples
92.8 billion
possible triples (N² × R) to score if nothing were sampled
The graph itself holds 2.9e-4% of them.
So the triples generator samples candidates, weighting entities by their clustering coefficient to explore sparse regions, where true links are more likely to be missing.
A use case: finding phage therapies
Phages are viruses that infect bacteria, and a promising answer to antimicrobial resistance. Their genomes are diverse and poorly understood, so deciding which lab experiments to run is hard. Link prediction can propose the hypotheses. Here is the paper's small phage graph with its five expert rules, running live. Step through:
Expert rules
- R1 host(Y) :- infects(X, Y).
- R2 broad_host_range(X) :- infects(X, Y1), infects(X, Y2), genus(Y1), genus(Y2), notEqual(Y1, Y2).
- R3 lysogeny(X, Y) :- phage(X), infects(X, Y), encodes(X, integrase), encodes(X, repressor).
- R4 anti_biofilm(X) :- encodes(X, depolymerase).
- R5 therapy_candidate_for(X, Y) :- lysogeny(X, Y), broad_host_range(X), anti_biofilm(X).
Nothing derived yet
Asserted facts only. Step through the stages.
Results
On a synthetic benchmark (LUBM) and a DBpedia subgraph, both with RDFS entailment rules, Poderoso almost always improves the embedding model it wraps. The best model differs per dataset (DistMult on LUBM, ComplEx on DBPedia20k), which is exactly why being able to plug in any model matters.
Reasoning adds little training time: at most about 34% (DistMult on LUBM), and for TransE the inferred triples even make training converge faster. Both KALE and pLogicNet are slower to converge.
Open challenges
Scale
KGE training and reasoning both grow with the graph and the rules; distributed training and reasoning are the way forward.
Uncertainty
KGE outputs are probabilistic, but probabilistic reasoners hurt scalability further.
Orchestration
Running the two engines naively in sequence can blow up training time.
Asking the right questions
Extracting plausible facts from a model without test data is still open; sampling strategies need work.
Optimisation
Choosing the model, fact-extraction strategy and reasoner per task calls for an optimiser, like a query optimiser.
Earlier work
The idea of loosely coupling knowledge graph embeddings with ontology-based reasoning first appeared in a 2022 preprint with Zoi Kaoudi and Volker Markl. The work was carried out during a master's thesis at TU Berlin.
See the 2022 preprint →Cite
@inproceedings{martinezlorenzo2025poderoso,
title = "Modular Neuro-Symbolic Knowledge Graph Completion",
author = "Martinez Lorenzo, Abelardo Carlos and
Perfilyev, Alexander and
Markl, Volker and
Clokie, Martha and
Sicheritz-Ponten, Thomas and
Kaoudi, Zoi",
booktitle = "VLDB 2025 Workshop: New Ideas for Large-Scale Neurosymbolic Learning Systems (LS-NSL)",
year = "2025",
url = "https://www.vldb.org/2025/Workshops/VLDB-Workshops-2025/LS-NSL/LS-NSL25_1.pdf"
}Graph and rules from Figure 3; accuracy from Table 2.