acceptodds
Withdrawn from ICLR 2027

Learning Transferable Topological Meta-Knowledge via Subgraph Embedding for Inductive Link Prediction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Inductive link prediction over knowledge graphs aims to generalize relational reasoning to entities that are entirely unseen during training. Existing methods alleviate the identity-binding limitation of conventional knowledge graph models, but remain challenged by cross-graph structural variation and the propagation of query-irrelevant information. To address these issues, we propose TOME (TOpological Meta-knowledge Extractor), a topology-based framework for inductive link prediction. TOME couples two complementary mechanisms: episodic meta-training over localized subgraphs, which aligns training with entity-disjoint graph transfer, and query-conditioned in-situ message passing, which incorporates the query relation into both message formulation and attention before neighborhood aggregation. This coupling enables TOME to learn transferable structural reasoning while suppressing irrelevant signals during message construction, without relying on target-entity-specific embeddings. Extensive experiments on WN18RR, FB15k-237, and NELL-995 demonstrate that TOME achieves strong overall performance in both inductive and transductive link prediction, outperforming topology-based baselines on most inductive and all transductive dataset–metric combinations. Ablation studies show that episodic meta-training and query-conditioned message passing make complementary contributions, with message-level query conditioning providing a substantial improvement. Further sparsity analyses support the robustness of TOME, while a qualitative case study illustrates its interpretable structural reasoning. These results demonstrate that coupling transfer-oriented meta-training with query-conditioned in-situ reasoning is an effective approach to entity-independent link prediction over unseen knowledge graphs.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
knowledge graph, knowledge graph reasoning, graph neural network, meta-learning, inductive link prediction

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_D5RaOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes TOME for inductive link prediction on knowledge graphs. TOME encodes the neighborhood of the query entity with recursive message passing and is trained with episodic meta-training. The query relation embedding is added to every message and to the attention input. The authors present this "pre-aggregation query conditioning" as a contribution and argue that it suppresses query-irrelevant signals during propagation. Experiments cover the 12 inductive splits of WN18RR, FB15k-237, and NELL-995, together with their transductive setting.

Critical strengths

The WN18RR results are notably higher than the competition. On the four inductive WN18RR splits, TOME's MRR is between 3.8 and 14.7 points higher than the best baseline (Table 1).

Critical weaknesses
  1. The premise that "query-irrelevant" information should be suppressed is not substantiated. The paper motivates the method by the need to suppress "irrelevant signals" and "topological noise" during propagation. This is a claim, not an established fact, and the paper should define what counts as irrelevant and show evidence that suppressing it helps. Information that is not directly relevant to the query can still give useful context about the entities involved, so propagating it is not harmful in general.

  2. The novelty of the contributions is unclear, and related work is missing or imprecise. The "Pre-Aggregation Query Conditioning" contribution claims that injecting the query relation into message construction is new. Query-conditioned message passing is well established: e.g., NBFNet , ULTRA, and other prior query-conditioning approaches are not discussed as such. Similarly, related work on meta-training are not well described, and it is unclear how the proposed episodic meta-training differs from it.

  3. The source of the improvements is unclear. The paper's own ablation (Table 5, two splits, single run) suggests that most of the gain comes from the episodic meta-training rather than from the query-conditioned messages. Applying the meta-training to plain RED-GNN raises WN18RR V1 MRR from 0.693 to 0.807, close to full TOME (0.836). Without meta-training, TOME (0.689) is no better than RED-GNN (0.693). Since it is unclear what in this meta-training is new (weakness 2), it is not possible to say which new element produces the gains.

Questions for authors
  1. How does an existing query-conditioned model perform when trained with existing meta-training approaches on the inductive splits? How does this compare with TOME?
  2. What exactly in the episodic meta-training differs from existing meta training approaches?
  3. Can the ablation in Table 5 be extended to all 12 splits, with results over multiple seeds?
Reviewer_DLG4OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

pos:

  1. Simple query-aware scoring with transferable structural training -> strong results, especially on WN18RR.

  2. Comprehensive experiments, including inductive/transductive settings, ablations, and graph sparsity analysis.

neg:

  1. "RED-GNN + In-situ Msg" and "TOME w/o Meta" seem similar but show different results.

  2. Numerical inconsistencies: MRR is lower than Hits@1 for GraIL on FB15k-237 V1.

Critical strengths
  1. The method combines a simple query-aware scorer with transferable structural training, and the reported results are strong, particularly WN18RR. Entity-independent modeling fits the inductive setting, where representations tied to training entities cannot simply be reused for unseen ones.

  2. The evaluation covers inductive and transductive settings, component ablations, and both source- and target-graph sparsity. This provides useful views of the proposed TOME approach.

Critical weaknesses
  1. The ablations do not clearly establish the incremental architectural contribution. For example, “RED-GNN + In-situ Msg” and “TOME w/o Meta” appear similar from their descriptions but obtain different results. The distinction, together with the episodic sampling specification, needs clarification for attribution and reproducibility.

  2. There are concrete numerical inconsistencies in Table 4 (GraIL's FB15k-237 V1 MRR as 0.252 but Hits@1 as 39.5%), because MRR must be at least Hits@1 under the same ranking protocol.

Questions for authors
  1. Can the authors reconcile the inconsistent metrics? Please also clarify whether the compared numbers use same candidate sets and filtering rules.

  2. What exactly differs between the two ablations I mentioned above, and how are episodic subgraphs and support/query splits sampled?

Reviewer_CSK5OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses inductive knowledge graph link prediction to predict missing relations for unseen entities. The proposed method combines episodic training on sampled subgraphs with query-conditioned message passing to learn transferable structural patterns. Experiments on three knowledge graph benchmarks show strong performance in inductive and transductive settings, with ablation studies highlighting the contributions of episodic training and query-aware message construction.

Critical strengths
  1. The method effectively addresses prediction on unseen entities without learning target-entity-specific embeddings or requiring target-side fine-tuning.

  2. TOME achieves overall better results in inductive settings and the best results in transductive comparisons compared with the baselines.

  3. The incorporation of episodic subgraph training is a promising approach to improving structural transferability. The reported results suggest that this training strategy can substantially improve inductive link prediction performance over conventional training.

Critical weaknesses
  1. Limited methodological novelty. TOME largely reuses RED-GNN's [1] recursive r-digraph expansion, query-conditioned attention, aggregation, scoring, and inductive zero-shot inference. Its primary additions are an additive query-relation vector in messages and episodic subgraph training. This raises significant concerns regarding the novelty of this work.

  2. Incomplete isolation of the proposed contributions. The paper provides useful ablations, but RED-GNN + Meta already obtains results close to TOME on WN18RR/FB15k-237. A matched factorial comparison is missing, and subgraph sampling and entity re-indexing are not isolated. Consequently, the claimed coupling and distinct meta-learning mechanism remain insufficiently established.

  3. Insufficient scalability evidence. Appendix B gives an asymptotic inference bound, but no matched wall-clock training/inference times, peak memory, or parameter counts. Tests on substantially larger graphs, such as YAGO3-10, are absent. Although dimensions 16–64 are searched, accuracy–efficiency tradeoffs at individual embedding dimensions are not reported or compared with baselines.

  4. Reproducibility details are incomplete. No TOME code-release link is provided in the supplied manuscript, and the selected per-split hyperparameter values are absent. Crucial episodic-task details—including subgraph sampling, task counts and sizes, and support/query partition ratios—are also unspecified. These omissions make it difficult to independently reproduce the principal claimed contribution.

  5. Statistical and reporting inconsistencies. Appendix G claims 138 wins in 144 comparisons, while the reported MRR/Hits@1 tables yield 128 strict wins against the three named baselines. The number of seeds or variance estimates is missing, preventing assessment of statistical significance. The FB15k-237 V1 depth study also reports an MRR of 0.597 across five layers, compared with 0.572 in the main tables.

  6. Limited evidence for broad structural transfer. All inductive splits come from only three dataset families. Robustness and component studies cover only selected splits, thereby limiting claims about generalization.

[1] Zhang, Yongqi, and Quanming Yao. "Knowledge graph reasoning with relational digraph." Proceedings of the ACM Web Conference 2022. 2022.

Questions for authors
  1. Can the authors run a parameter-matched factorial ablation of episodic training and query-conditioned message content on the same RED-GNN backbone? Can they separately test subgraph sampling, local candidate normalization, and entity re-indexing, and explain the difference between RED-GNN and TOME without Meta?

  2. Can the authors compare TOME and RED-GNN on larger graphs (e.g., YAGO3-10), reporting training time, inference latency, peak memory, and parameter counts? Please also report MRR and runtime across the searched embedding dimensions, especially d=16d=16 and d=32d=32.

  3. Is the implementation publicly available, or will an anonymized code artifact be supplied? Can the authors release the final configurations, the exact subgraph/meta-task sampling procedure, the support/query ratios, and the random seeds to enable reproducibility?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.