acceptodds
Withdrawn from ICLR 2027

RuleTopo: Neuro-Symbolic Synergy for Inductive Knowledge Graph Completion

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Knowledge Graph Completion (KGC) infers missing links to enrich incomplete knowledge graphs, categorized into transductive and inductive scenarios. Inductive KGC poses significant challenges, particularly in fully inductive settings with entirely new relations, and still lacks effective modeling paradigms. Currently, rule-based approaches rely on known relation semantics and struggle to generalize to new relations, while relation-graph methods mainly model statistical co-occurrence structures and overlook transferable logical constraints between relations, thus performing poorly when entities and relations are completely unseen. To tackle these issues, we propose RuleTopo, a novel Graph Neural Network-based approach that fuses rule learning with a global topology relation graph, supporting fully, semi-, and traditional inductive KGC. Our approach extracts Horn rules and maps them into continuous embeddings, explicitly integrating symbolic rules into the attention mechanism of a relational graph neural network to enable rule-guided topological message passing. Furthermore, we introduce a rule-consistency loss to constrain the logical structure of relation embeddings during training continuously. Experiments show that on the -100/-50/-0 inductive splits of the FB15K-237 and NELL995 datasets, RuleTopo demonstrates higher efficiency and interpretability. This approach demonstrates a scalable, highly generalizable neuro-symbolic cooperative paradigm, offering a new direction for inductive KGC.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Neural Network, Knowledge Graph Completion, Inductive Learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_6b8rOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies inductive knowledge graph completion, focusing on the fully inductive setting in which neither entities nor relations are shared between the training and test graphs. The authors propose RuleTopo, which encodes Horn rules mined by AMIE+ into relation-level descriptors and incorporates them into a relation-graph attention network with six types of topological patterns. The resulting relation representations guide entity-level message passing and triple scoring. The method also introduces a rule-consistency loss that encourages relation compositions derived from the mined rules to satisfy additive constraints in the embedding space.

The main contribution is the integration of symbolic rules and global relation topology during relation representation learning, rather than applying rules only after prediction.

The authors evaluate the method through baseline comparisons and component ablations on several inductive splits of FB15K-237 and NELL995. The reported results show improvements over InGram and TARGI in the fully inductive setting, although these improvements do not extend to all other settings. There are unexplained discrepancies between the main results and the ablation results, and the rule encoding procedure and test-time pipeline are insufficiently specified. Consequently, the experiments provide preliminary support for the approach but do not yet establish the effectiveness of its core mechanism or the robustness of its advantages.

Critical strengths
  1. A well-motivated mechanism for relation-level integration

Fully inductive knowledge graph completion requires generalization to unseen relations, not merely unseen entities. The proposed method incorporates rule descriptors directly into relation-graph attention, allowing symbolic information to influence representation learning rather than functioning solely as a separate post-processing module.

This constitutes a concrete methodological contribution. The relation level is a sensible interface for combining graph structure with relational composition patterns, and the design offers potential for further extension.

However, I regard this as a plausible and concrete modeling approach, rather than an established mechanism for general-purpose logical transfer.

  1. Promising reported results in the most relevant fully inductive settings

Compared with TARGI, RuleTopo improves MRR from 0.251 to 0.272 on FB-100 and from 0.315 to 0.334 on NL-100, corresponding to relative improvements of approximately 8.4% and 6.0%, respectively.

These results directly address the challenging setting emphasized by the paper and use a strong topology-aware relation-graph method as the comparator. If the experimental settings are consistent, the results are reproducible, and the gains are stable, these findings would provide the main empirical basis for recommending acceptance.

At present, however, the absence of statistics across multiple runs and the discrepancies between the reported tables make this strength conditional.

Critical weaknesses
  1. The core mechanism for generalization to unseen relations is insufficiently specified

The paper argues that rules help the model handle entirely unseen relations. However, being independent of entity identities does not, by itself, imply that a rule representation can transfer to an unseen relation vocabulary.

The current manuscript does not sufficiently specify:

  • How a consistent rule-feature space is constructed across graphs;
  • Whether rule encoding depends on specific relation identifiers;
  • How rules and representations for unseen relations are constructed in the test graph.

Furthermore, the rule signal used in attention is formed through symmetric summation, while the consistency loss models composition through commutative vector addition. Horn rules and relational path compositions generally contain directional and order-sensitive information. The manuscript does not adequately explain which logical information is preserved by these representations.

These issues bear directly on the central claim: whether the model transfers logical structure or instead benefits from additional statistical descriptors of the target graph. Without a sufficiently precise operational specification, the method cannot be reliably reproduced, and it is difficult to assess whether it achieves the claimed neuro-symbolic synergy.

  1. The experimental evidence is insufficient to support the strength of the conclusions

The most immediate concern is the inconsistency in the reported performance of the full model on FB-100:

MetricMain Results: Table 1Ablation Results: Table 3
MRR0.2720.253
Hits@100.4330.429
Hits@10.1940.172

The paper does not explain these discrepancies. If different settings were used, the ablation results do not directly account for the main results. If the settings were identical, the reporting needs to be corrected and the source of the discrepancies explained.

In addition, the results on NL-50 and NL-0 do not support the broad claim that the method consistently outperforms strong baselines. The manuscript also lacks variability estimates across multiple runs, controls using randomized or shuffled rule information, and empirical comparisons with ULTRA and AdaRPT, both of which are discussed in the related work.

The paper’s contribution rests primarily on empirical validation. Experimental consistency and appropriate controls are therefore essential to establishing the effectiveness of the proposed mechanism, rather than merely matters of presentation. The current evidence does not adequately distinguish stable methodological gains from differences in experimental settings or random variation.

Questions for authors

Question 1: Please specify the complete rule encoding procedure and test-time pipeline

Is AMIE+ run again on the test support graph? What are the exact inputs, encoding procedure, and parameterization of frulef_{\mathrm{rule}}? How are entirely unseen relations handled, and how is feature construction made independent of training-specific relation identifiers? Does the encoding preserve rule-head versus rule-body roles, atom order, and variable-sharing structure? How should the descriptions of rule embeddings in the main text and appendix be reconciled?

Question 2: Please explain the discrepancies between the main results and the ablation results

Were the two sets of full-model results on FB-100 obtained using the same data, model configuration, checkpoint selection procedure, and evaluation protocol? If not, what motivated the different settings? Do the gains attributed to the individual components remain when evaluated under a consistent experimental setup?

Question 3: What evidence shows that the gains arise from rule structure rather than additional features or random variation?

Under matched training settings, do genuine rules outperform shuffled rules or non-logical features with comparable statistical properties? Are the improvements stable across multiple runs? Does the proposed mechanism retain an advantage when compared with relevant fully inductive methods?

Reviewer_ExZXOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes RuleTopo, which employs AMIE+ to mine Horn rules and encodes these rules as relation-level features. These features are then integrated into a relational graph attention mechanism, structured according to six distinct topological patterns, while a rule-consistency loss is applied to constrain the relational representations. The core approach leverages explicit rules to augment the topology of relation co-occurrences, thereby enabling fully inductive knowledge graph completion, even for entities and relations that have not been encountered during training.

Critical strengths

The research problem is significant. The paper focuses on unseen relations in a fully inductive setting, which is more challenging than dealing solely with unseen entities.

Critical weaknesses
  1. The current ablation studies fail to conclusively demonstrate that the performance gains stem from the logical rules themselves. Removing the rule injection alters both the input information and the model capacity. Therefore, it cannot be ruled out that the observed gains are primarily derived from additional statistical features or parameters.

  2. The experimental results contradict the claims made in the abstract. In Table 2, RuleTopo achieves an MRR of 0.234 on the NL-0 dataset—a figure lower than those of NBFNet, TARGI, and InGram. Its Hits@1 performance also falls short of several baselines. Consequently, the assertion in the abstract and conclusion that the model "consistently outperforms strong baselines" is an overstatement.

  3. There are direct inconsistencies in the experimental results, which complicate the assessment of performance improvements and the validity of the ablation study conclusions. The results for the same complete model on the FB-100 dataset differ between Table 1 and Table 3. Specifically, the values ​​for MRR, Hits@10, and Hits@1 are inconsistent: Table 1 reports RuleTopo as 0.272 | 0.433 | 0.194, whereas Table 3 reports 0.253 | 0.429 | 0.172.

  4. The primary contributions of this paper should be the injection of rule-based features and the application of rule-consistency regularization; the overall relational graph architecture itself cannot be claimed as a novel contribution.

Questions for authors
  1. Explain the differences between the full models presented in Table 1 and Table 3, verify any anomalous metric combinations, and clarify the origin of the ablation results (0.253, 0.429, 0.172) on the FB_100 dataset, which are identical to those reported for TARGI [1].

[1] Ding, Ling, et al. "Towards global-topology relation graph for inductive knowledge graph completion." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 39, No. 11, 2025.

  1. Since ULTRA is already cited in this paper, it should be included in the experimental comparisons. Please provide a comparison under comparable training conditions, or explicitly explain any differences in the experimental protocols.

  2. In Table 2, RuleTopo achieves an MRR of 0.234 on the NL-0 dataset—a figure lower than those of NBFNet, TARGI, and InGram. Its Hits@1 performance also falls short of several baselines. Consequently, the assertion in the abstract and conclusion that the model "consistently outperforms strong baselines" is an overstatement.

Reviewer_WwcUOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces RuleTopo, a neuro-symbolic framework for fully inductive KGC where both entities and relations can be entirely unseen at test time. By combining AMIE+-extracted Horn rules with a global relation topology graph, the authors project pre-mined symbolic rule embeddings into a Multi-Head Attention GCN and enforce logical structure via a custom rule-consistency loss.

Critical strengths
  1. Handling unseen relations in fully inductive KGC is notoriously tricky because learned relation embeddings cannot simply be looked up.

  2. Instead of forcing heavy, differentiable rule-mining layers end-to-end, the approach treats static rule embeddings as soft priors within GATv2 attention mechanisms.

  3. The empirical evaluation is thorough, outperforming strong baselines like InGram and TARGI on FB-100/NL-100 splits.

Critical weaknesses
  1. Section 3.2 states that if no rules are mined for a relation rr, its rule embedding defaults to zero (ur=0u_r = 0). In fully inductive settings, where relation co-occurrence can be extremely sparse, how often does rule extraction fail?

  2. AMIE+ often mines low-confidence or noisy rules in dense graphs. Does low-quality rule injection distort the topology representation learned by the relation GNN? An analysis showing how model performance degrades as rule confidence thresholds vary would clarify the model's robustness.

  3. The paper claims higher efficiency in both the Abstract and Introduction, but lacks any quantitative wall-clock time, GPU memory usage, or inference latency benchmarks. Given that the pipeline requires AMIE+ pre-processing plus multi-bin relation-graph attention, empirical proof is needed to back up the efficiency claim against baselines like InGram and NBFNet.

  4. Missing an important baseline LLM4EA.

Questions for authors

Please see above.

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

Related papers

Discussion (0)

Sign in to comment.