PaperlayerView exact revision on arXiv

Source-linked AI summary

Reification as a Transferable Vocabulary: Zero-Shot Link Prediction with Vanilla GNNs

Camille Pradel

arXiv:2609.11347v1cs.LGcs.AI

TL;DR

The paper asks whether zero-shot transfer across unseen knowledge graphs requires a dedicated foundation-model architecture or can arise from a shared representation. It reifies facts into homogeneous graphs with anonymous relation nodes, trains vanilla GNNs on one knowledge graph, and finds that the best GAT approaches ULTRA across inductive benchmarks while the same vocabulary transfers preliminarily to databases.

  • Problem

    Existing zero-shot knowledge-graph transfer relies on architectures with dataset-specific relation or entity vocabularies, while relational foundation models use mechanisms unavailable on bare knowledge graphs.

  • Method

    The paper reifies each fact as a node linked to its subject, object, and relation type through six fixed meta-relations, enabling generic node-inductive GNNs to share one finite vocabulary.

  • Results

    Five textbook GNNs transfer to 40 inductive benchmarks, with GAT scoring 0.3600 ± 0.0055 against ULTRA’s 0.3617 on 13 benchmark families unseen by both pretraining corpora.

  • Takeaways & Limitations

    A fixed structural representation can carry most of the demonstrated zero-shot inductive transfer without dedicated foundation-model machinery and can extend from knowledge graphs to relational databases.

  • Takeaways & Limitations

    GAT trails ULTRA by 0.017 on the full 40-benchmark aggregate, and the remaining gap is not closed by choosing a different single training graph.

Abstract

from arXiv · show

Knowledge graph foundation models such as ULTRA achieve zero-shot link prediction on unseen graphs through dedicated architectures that hard-code a transfer mechanism. In this work we move that mechanism out of the architecture and into the representation, by \emph{reifying} the input graph: every fact becomes a node, connected to its subject, object, and relation type through a fixed vocabulary of six meta-relations, with relation types as anonymous shared nodes rather than model parameters. On this representation, five textbook GNNs (GAT, GINE with sum and with mean+max aggregation, GraphSAGE, R-GCN), each trained on a single knowledge graph of 4,245 triples for 30 minutes on one NVIDIA A100, transfer zero-shot to 40 inductive link-prediction benchmarks. The best of them, an off-the-shelf GAT, matches ULTRA, a dedicated foundation model pretrained on three graphs, across ULTRA's own evaluation suite. The same fixed vocabulary extends to relational databases, a row becoming an entity and a foreign-key column a relation type; a preliminary probe on two unseen databases, with no cell values, schema text or in-context labels, shows a model of this family pretrained on three knowledge graphs ranking foreign-key targets far above random-initialization and degree controls. We release the code, the checkpoints, and the evaluation pipeline for all 40 benchmarks.

1 Introduction

The paper moves zero-shot transfer from specialized model architectures into a shared reified representation, then tests whether generic GNNs can transfer across unseen graphs and databases. Five textbook GNNs transfer to 40 inductive benchmarks, while the best GAT approaches ULTRA despite substantially simpler machinery.

  • Contribution: Reification moves transfer from the architecture into a fixed representation where facts become nodes and relation types become anonymous shared nodes.The representation uses a universal structural vocabulary rather than dataset-specific entity or relation parameters.
  • Results: Five textbook GNNs trained on one 4,245-triple knowledge graph transfer zero-shot to 40 inductive benchmarks under one frozen recipe.The recipe uses 64 dimensions, 12 layers, 30 minutes on one NVIDIA A100, and three fixed seeds.
  • Results: Over all 40 benchmarks, GAT averages 0.3565 ± 0.0049 against 0.3732 for ULTRA, while scoring 0.5457 against 0.5224 on twelve inductive-(e) GraIL benchmarks.On 13 benchmark families absent from both pretraining corpora, GAT scores 0.3600 ± 0.0055 against ULTRA’s 0.3617.
  • Database probe: A preliminary database probe applies the same representation to two unseen databases without cell values, schema text, or in-context labels.The supplied passage introduces this probe but does not report its numerical outcome.
  • Scope: The paper does not claim to beat the state of the art: ULTRA remains ahead on the full aggregate, and later KGFMs publish higher inductive-(e) figures.The authors identify the contribution as demonstrating substantial transfer from a fixed structural representation without dedicated machinery.

2 Related work

Related work frames transfer as relying on structural invariants, schema or relation-name text, and in-context conditioning, with existing systems typically specializing in one data regime. This paper instead makes the rewritten data itself the transferable substrate for both knowledge graphs and relational databases.

  • Knowledge-graph foundation models: Existing KG foundation models use dedicated architectures whose transfer mechanism remains in the model, whereas this paper puts the mechanism in the data.The reified graph can be read by any node-inductive GNN rather than only by a specialized relation-graph architecture.
  • Relational foundation models: Relational foundation models generalize to unseen schemas through schema text and in-context conditioning, but they are not designed to cover standard KG link prediction.Their transfer channels include cell text, column names, and labeled target context, which bare KG benchmarks do not provide.
  • Transfer mechanisms: The paper organizes transfer mechanisms into structural invariants, schema or relation-name text, and in-context conditioning on labeled examples.Its representation uses structural invariants alone at inference.
  • Reification: Reification builds on classical edge-to-node transformations and prior edge-centric link-prediction methods, but here serves as a universal representational substrate rather than an architectural device.The distinction is that the data itself is rewritten into a shared vocabulary consumable by generic node-inductive GNNs.
  • Expressiveness and equivariance: The method structurally realizes entity×relation equivariance because permuting relation types permutes anonymous relation nodes.This connects reification to work on unseen-relation prediction and double equivariance.

3 Method

The method reifies each fact into a node-centered homogeneous graph with a fixed meta-relation vocabulary, then extends the same construction to database rows, foreign-key columns, and tables. Query conditioning lets generic inductive GNNs perform link prediction without dataset-specific parameters.

  • Reified representation: Each fact becomes a FACT node connected to its subject, object, and relation-type nodes through six fixed directed meta-relations.The six edges are has_subject, subject_of, has_object, object_of, has_type, and has_instance.
  • Reified representation: The universal graph vocabulary contains three node kinds and six meta-relations, with relation types represented as anonymous shared nodes rather than learned identities.Sharing a relation node makes same-kind structure explicit in the graph instead of storing it in an embedding table.
  • Relational databases: A database row becomes an entity, each foreign-key cell becomes a FACT node, each foreign-key column becomes a REL_TYPE node, and each table becomes a CLASS node.The database extension adds the is_a and has_member meta-relation pair while preserving the shared vocabulary.
  • Structural properties: The representation preserves relation direction structurally and makes a homogeneous GNN automatically equivariant to relation-type permutations.A two-hop path through a fact node corresponds to one typed message-passing hop in the original graph.
  • Query conditioning: Query conditioning initializes nodes from their kinds plus markers for the known endpoint role and queried relation, because unconditioned node invariants cannot rank links.There are no per-relation, per-entity, or per-dataset parameters.
  • The five backbones: The five backbones are off-the-shelf GAT, GINE variants, GraphSAGE, and R-GCN models sharing a 64-dimensional, 12-layer configuration.Reification roughly multiplies node count by 20 and doubles graph diameter, motivating the depth choice.

4 Experimental setup

The evaluation uses a controlled comparison of models trained on one knowledge graph against ULTRA across 40 zero-shot inductive benchmarks, while excluding one prohibitively expensive split.

  • The training data comprise one FB15k-237-Inductive v1 graph with 1,594 entities, 180 relations, and 4,245 triples.
  • 40 inductive benchmarks are evaluated with filtered ranking and absolute MRR, covering GraIL, InGram, and extended benchmark families.
  • ULTRA is compared using its released three-graph checkpoint, with a controlled re-run evaluated on identical splits, candidate sets, and tie-breaking.
  • The controlled comparison reports that the all-40 gap is 0.017 MRR, whereas comparison with ULTRA’s published figures widens the gap to 0.025.
  • The original 41-benchmark suite omits HM:indigo because one evaluation ran 4h38 without finishing and the exclusion shifts aggregates by about +0.003 MRR.
  • The 30-minute training cap binds only for GAT, and extending it to 20 epochs adds +0.003 MRR on the full evaluation.

5 Results

Across 40 unseen graphs, vanilla GNNs transfer reliably on the reified representation, with GAT matching ULTRA on a symmetric subset but trailing on the full aggregate. The main weakness is isolated entities, while changing the single training graph does not explain the remaining gap.

  • 5 Results: All five textbook GNNs transferred to 40 unseen graphs; R-GCN, the weakest, still averaged 0.2350, and the model ordering was stable apart from one swap.The reported ordering was GAT > GINE-mean+max > GINE-sum > GraphSAGE > R-GCN.
  • 5 Results: A 16-draw GAT portfolio averaged 0.4293 ± 0.0047 on the 25-benchmark GraIL+InGram suite versus 0.4299 for ULTRA, matching it in the single-small-KG regime.On the full 40-benchmark aggregate, GAT was 0.0167 behind ULTRA.
  • 5.2 Where the gap lives: families and regimes: The GAT was ahead of ULTRA on WordNet and NELL, tied on Freebase, and behind on Wikidata and the Hetionet-derived splits.The per-family comparison is relative to each model’s different pretraining corpora.
  • 5.2 Where the gap lives: families and regimes: The 13-benchmark symmetric comparison found the two models indistinguishable, but the primary all-40 aggregate still had ULTRA ahead by 0.017.The symmetric subset contains nine NELL splits, three Hetionet-derived splits, and one Metafam split.
  • 5.3 The one failure mode: isolated entities: On FBNELL, all fifteen models collapsed to MRR 0.076–0.106 with Hits@1 exactly 0 because five isolated entities receive no propagated messages.FBNELL is the only one of the 40 splits whose test inference graph contains isolated entities.
  • 5.4 Does the choice of training graph matter?: Changing the single training graph did not explain the transfer profile: CoDEx-Small lost 0.008 on Wikidata-family benchmarks and 0.061 on the 40-graph mean.Larger ConceptNet100k and AristoV4 candidates also failed to help on weak families, although their sampled-subgraph results were not comparable to the main results.

6 The same vocabulary reads a relational database

The shared reified vocabulary transfers from knowledge graphs to foreign-key graphs, even without cell values, schema text, or in-context labels. Performance is strongest when structural patterns distinguish targets, while interchangeable targets remain difficult.

  • Probe results: Hits@10 reaches 0.78–1.00 on four Formula-1 driver and constructor keys, versus 0.00 for random initialization.The transferring model also exceeds a degree heuristic on these keys.
  • Probe results: Across both databases and pretraining seeds, keys targeting near-interchangeable races or posts remain near random.The bare foreign-key structure provides no distinguishing signal for those targets.
  • Comparison with ULTRA: ULTRA-50g exceeds both controls on five keys per database, while the generic model leads on most driver, constructor, and author keys.The two models split advantages across key types and reach parity on post keys for comments and edit history.
  • Corpus dependence: The corpus matters: one-graph pretraining transfers poorly, whereas three transductive knowledge graphs support transfer across the tested database keys.The one-graph model transfers on none of the rel-f1 entity-referencing keys, while the three-graph setting transfers on all ten strong keys.
  • Scope: The database experiment remains a two-database probe using one reified model family and one external KG foundation model, without a supervised specialist comparison.The authors frame it as evidence for the shared vocabulary rather than general relational-database capability.

7 Limitations

The method’s remaining limitations concern transfer scope, isolated entities, computational cost, and the narrow evidence supplied by the database probe. These constraints leave both architectural and scaling questions open.

  • Transfer scope: ULTRA remains ahead by 0.017 overall, and the demonstrated single-graph transfer is limited to Freebase, WordNet, and NELL-derived families.Later KG foundation models also publish higher figures on the inductive-(e) splits.
  • Isolated entities: Plain-readout models struggle with isolated entities, while the residual gap on that split remains 0.43 versus 0.47 for the best of six models.A controlled readout ablation is still needed to separate readout effects from training-graph effects.
  • Computational cost: Reification multiplies node count by roughly ×20, doubles graph diameter, and reaches memory limits on 100k-fact training candidates.The reported 30-minute cost therefore applies to the demonstrated graph-size regime, not a scaling law.
  • Scope: All results concern structural link prediction on featureless graphs, and the database probe does not establish general relational-database capability.The entity-type nodes remain untested because joint KG-and-database pretraining has not been performed.

8 Conclusion

The paper argues that zero-shot inductive link prediction can rely on a shared structural representation rather than a dedicated foundation-model architecture. Reification supplies that vocabulary and transfers across knowledge-graph benchmarks and a preliminary database probe.

  • Conclusion: An off-the-shelf GAT reaches 0.3600 versus ULTRA’s 0.3617 on 13 benchmarks whose families neither model saw during pretraining.Across all 40 benchmarks, it trails ULTRA by 0.017, with more than half the gap concentrated in one isolated-entity split.
  • Representation: Reification represents datasets with three node kinds and six meta-relations, extending to four node kinds and eight meta-relations for databases.Five textbook GNNs trained on one 4,245-triple graph transfer to 40 inductive benchmarks under this vocabulary.
  • Database probe: A three-graph-pretrained model ranks foreign-key targets above random-initialization and degree controls on two unseen databases.The result supports the vocabulary’s cross-data-model transfer in a preliminary probe.

A Protocol details

The experiments use a frozen training recipe, fixed hardware, and controlled checkpoint selection across the 15 KG models. Reproducibility checks verify both stored configurations and representative evaluation and training paths.

  • Frozen recipe: All 15 models use dimension 64, 12 layers, dropout 0.2, a 1,800-second cap, and the FB15k-237-Inductive v1 training split.Only the backbone and seed vary, with Tier-1 selection based on the training graph’s validation MRR.
  • Hardware: Every reported number uses one NVIDIA A100 with one job at a time, including the ULTRA rerun.The paper’s timing claims all refer to this device.
  • Code non-drift control: Four commits between training batches prompted a pre-specified non-drift verification before the remaining seeds were trained.The acceptance threshold was fixed at 0.005 MRR.
  • Reproducibility: Re-evaluation reproduced five stored checkpoint results within 0.0009, while retraining an R-GCN configuration matched validation MRR within 0.0003.The checks cover both evaluation and training paths.
  • Configuration audit: The 15 stored configurations were audited field by field, including absent later-introduced flags treated as protocol defaults.Treating absent keys as differences would create eleven spurious deviations.

B Transparency notes

The appendices document evaluation decisions, provenance, exclusions, and reporting conventions, including safeguards around anomalous splits and training regimes.

  • The excluded split changes the 41-benchmark flat-mean ULTRA aggregate from 0.3747 to 0.3732 and the reference model from 0.3359 to 0.3366.
  • Removing FBNELL’s five isolated entities would raise every measured cell by 0.23 to 0.39 MRR, so the authors retain the original candidate pool.
  • Sampled-subgraph training lowers the reference average to 0.2756 versus 0.3565 for full-graph GAT, making those regimes non-comparable to the full-graph results.
  • The database probe uses three passing pretraining seeds, while a fourth re-pretraining failed the fixed 95% selection-score sanity gate.
  • The 15 × 40 benchmark matrix records per-seed filtered MRR, while aggregate values are unweighted means with standard deviations across three seeds.

C Full result matrices

The appendix matrices expose per-seed and per-benchmark results, document alternative training graphs, and detail the FBNELL anomaly underlying the aggregate comparisons.

  • Sixteen frozen-recipe GAT trainings average 0.4293 ± 0.0047 on 25 GraIL and InGram benchmarks, versus 0.4299 for ULTRA-3g.
  • On FBNELL, all fifteen models have no query at ranks 1–4 and mode rank 6, while isolated-answer queries determine the anomaly’s structure.
  • Table C1 reports filtered MRR means and standard deviations for each backbone across 40 benchmarks, alongside locally rerun and published ULTRA-3g values.
  • The 15 × 40 matrix contains one filtered-MRR cell per backbone seed and benchmark, with accompanying Hits@1, Hits@3, and Hits@10 values.
  • Alternative training graphs are reported as per-family ΔMRR relative to a reference using the same training regime, with sampled-subgraph columns separated from full-graph comparisons.

D The KG→RDB probe: protocol and per-key results

The probe reifies relational databases using the same anonymous graph vocabulary and evaluates foreign-key recovery under controlled, filtered ranking conditions.

  • Protocol: Each database row becomes an entity, each foreign-key column an anonymous relation type, and each non-null foreign-key cell a fact with six meta-relation edges.
  • Scope: The probe excludes cell values, column names, table names, and timestamps, and the database graphs contain no class nodes despite the broader schema extension.
  • Protocol: For every foreign-key column, half the cells remain as support and the other half become subject-to-object queries, with referenced-table rows as candidates and known targets filtered.
  • Models: The probe’s three-KG model is a mean+max GINE backbone pretrained on FB15k-237, WN18RR, and CoDEx-Medium, with three seeds passing the sanity gate.
  • Results: The generic model pretrained on three knowledge graphs transfers on four rel-f1 driver and constructor keys and six rel-stack keys, while six one-graph models transfer on no entity-referencing rel-f1 key.
  • Results: Transfer is strongest for dense co-key contexts and structurally rich targets, while near-interchangeable event targets and badges.UserId remain near random across databases and seeds.
Loading 2609.11347v1…