acceptodds
Under review as a conference paper at ICLR 2027

Towards Hierarchical Invariant Semantics for Inductive Knowledge Graph Reasoning

Abstract

Inductive Knowledge Graph Reasoning (IKGR) aims to infer missing information for unseen entities and relations given an observed Knowledge Graph (KG). However, most existing GNN-based methods suffer from limited generalization and poor extrapolation due to indiscriminate message passing mechanism, which entangles task-relevant signals with spurious correlations confined to the training environment. In the inductive settings, such spurious patterns dilute invariant learning and significantly degrade performance. To address this challenge, we propose HiKGR, a novel framework that models Hierarchical invariant semantics for inductive Knowledge Graph Reasoning, thereby improving generalization and extrapolation. Specifically, we provide an analysis that the learned representation is prone to spurious associations without explicit invariant constraints. Then we design a multi-environment generator that simulates diverse KG domains, and enhance extrapolation by exploring invariant semantic factors at hierarchical relation, entity, and subgraph levels. Further, we develop a bi-level optimization strategy to jointly encourage cross-environment invariance while preserving domain diversity. Extensive experiments on real-world datasets demonstrate that the proposed HiKGR consistently outperforms state-of-the-art methods.

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

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_B4xPOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2(−1)
Soundness
2(+1)
Read review
Paper summary

The paper proposes HiKGR, a hierarchical invariant semantic learning framework for inductive knowledge graph reasoning with unseen entities. HiKGR aims to reduce reliance on environment-specific spurious correlations introduced by indiscriminate GNN message passing. Across three real-world benchmarks comprising 12 entity-disjoint inductive settings, HiKGR is reported to improve MRR, Hits@1, and Hits@10 over prior rule-based and message-passing-based methods. The paper has an interesting intuition, but its empirical positioning is currently incomplete, and its invariance claims are overstated. A stronger comparison with recent inductive KGC methods and a serious discussion or evaluation against foundation models are necessary before the claimed contribution can be properly assessed.

Critical strengths

S1 The paper addresses an important limitation of inductive knowledge graph reasoning: GNN-based models may capture environment-specific structural correlations that do not generalize to unseen entities.

S2 The hierarchical invariant learning mechanism provides a concrete way to model cross-environment semantics at the relation, entity-path, and query-subgraph levels.

S3 This paper is easy to follow.

Critical weaknesses

W1 The paper should more clearly distinguish HiKGR from existing graph augmentation, multi-environment learning, consistency regularization, and invariant-learning methods. At present, the proposed environment generator plus cross-environment consistency objective appears conceptually close to learning robustness against self-generated structural perturbations.

W2 The connection between the stated objective I(Y;S∣Z)I(Y; S∣Z) and the implemented relation-level, entity-level, and subgraph-level losses is unclear. The paper does not provide a derivation showing that these losses optimize a meaningful bound or surrogate of the claimed information-theoretic objective.

W3 The mask-density constraint may preserve graph size or sparsity, but it does not ensure that essential facts, relation semantics, connectivity patterns, or reasoning paths are preserved. In particular, dropping edges can easily eliminate the only evidence supporting a query, while retaining density alone does not prevent semantically implausible environments. This raises the concern that the model may learn invariance to arbitrary graph corruption rather than invariant KG semantics.

W4 The paper repeatedly refers to “spurious correlations,” “stable semantics,” and invariant knowledge, but no causal assumptions or real environment shifts are specified. Besides, the paper uses “generalization,” “extrapolation,” “inductive reasoning,” and OOD robustness in a somewhat interchangeable manner. These are not equivalent notions. The task seems primarily to be inductive knowledge graph completion/link prediction over unseen entities, rather than general KG reasoning or broad extrapolation.

W5 The comparison set does not adequately reflect recent work on transferable relation structures and cross-graph generalization. It is not sufficiently clear whether all baselines use identical inductive splits, candidate entity sets, filtered-ranking protocols, hyperparameter budgets, training epochs, and official implementations.

Questions for authors

Main question pls see weakness.

Q1 If only compared in the transductive setting, ULTRA cannot be regarded as direct support for the claim regarding "related work in Inductive knowledge graph reasoning".

Q2 The first sentence of the conclusion is obviously repetitive. And pls double check Table 5, dataset name and details are wrong.

Reviewer_cQSUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes HiKGR, a GNN-based framework for inductive knowledge graph reasoning that aims to improve generalization by learning hierarchical invariant semantics across multiple, synthetically generated environments. The method introduces a trainable multi-environment generator (via edge masks and Gumbel-Sigmoid) to induce diverse structural perturbations while preserving salient semantics, and enforces cross-environment consistency through three complementary regularizers at relation, entity, and subgraph levels within a bi-level optimization scheme. Experiments on WN18RR, FB15K-237, and NELL-995 show improvements over several recent inductive baselines.

Critical strengths

1: The paper combines trainable multi-environment generation with cross-environment semantic consistency at relation, entity, and subgraph levels. These components address different aspects of relational reasoning and provide a coherent approach to learning representations under structural variation. The ablation results generally support the usefulness of the proposed components, making this integration a meaningful contribution.

2: The paper evaluates HiKGR against several recent baselines across twelve inductive benchmark settings and three ranking metrics. The results are particularly strong on Hits@10, and the component ablations support the overall design. The theoretical propositions also provide useful motivation for studying spurious correlations and environment diversity, although they do not establish a direct guarantee for the implemented training objective.

Critical weaknesses

1: The theoretical results motivate invariant representation learning but do not directly establish that the proposed hierarchical regularizers control the conditional mutual information term in Proposition 1. These regularizers are acknowledged to be surrogate objectives rather than estimators or bounds of that term. Moreover, the generalization bound in Proposition 4 relies on a conditional coverage assumption whose applicability to the evaluated inductive settings is not empirically established. This weakens the connection between the theoretical analysis and the claimed generalization mechanism.

2: Although HiKGR achieves strong results in several settings, its improvements are not uniform: for example, it underperforms KnowFormer on NELL-995_v1 MRR and WN18RR_v3 Hits@1. The paper reports means over three runs without corresponding variability estimates. While the ablations demonstrate the usefulness of the proposed components, they do not fully disentangle the effects of environment-generation capacity, diversity optimization, and hierarchical consistency regularization.

Questions for authors

1: The proposed relation-, entity-, and subgraph-level regularizers encourage representation consistency but are not established as bounds or estimators of the conditional mutual information term. Could the authors clarify what formal or empirical evidence supports their effectiveness in controlling label-dependent environmental variation? Additionally, how realistic is the conditional coverage assumption of Eq. (38) for the evaluated inductive splits?

2: Table 2 compares HiKGR with general graph augmentation approaches, while Figure 3(b) and Table 8 provide component ablations. Could the authors further distinguish the effect of the learned environment generator from that of hierarchical consistency training under matched model capacity and training budgets? In particular, how does the mask-density constraint preserve task-relevant semantics rather than merely maintaining a desired number of edges?

3: HiKGR does not outperform the strongest baseline in every MRR or Hits@1 setting. Could the authors provide standard deviations or confidence intervals for the reported three-run results and analyze the settings where HiKGR underperforms?

2 comments

Thank you for the additional analyses. The new results address most of my empirical concerns. Some limitations remain regarding the theoretical justification and the assumptions underlying the generalization analysis. I therefore keep my original score of 2.

Thank you for the additional explanation and experiments. However, my main theoretical concern remains: the proposed regularizers are still only empirically connected to cross-environment consistency, and the assumptions underlying the generalization analysis are not directly verified.

Reviewer_GXc5OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes HiKGR, a model for inductive knowledge graph reasoning (IKGR), which aims to infer missing facts in knowledge graphs (KGs) containing unseen entities. To apply invariant learning (IL) to IKGR, HiKGR generates subgraphs with diverse structures from the observed KG through multi-environment construction. For each subgraph, the model updates both global relation representations and query-conditioned entity representations, ultimately using the entity vectors to make predictions. Empirically, HiKGR demonstrates performance that is competitive with or superior to existing baselines on IKGR benchmarks.

Critical strengths
  • The paper proposes a novel training methodology to adapt invariant learning to the IKGR task.
Critical weaknesses
  • This paper only addresses the entity-inductive setting; however, in real-world scenarios, relations can also be unseen, which has already been extensively studied in prior work [1, 2, 3]. Compared to the existing literature that handles both inductive entities and relations, focusing solely on the entity-inductive setting offers no clear advantage.

  • Beyond the application of invariant learning, the technical novelty of this work is unclear. Query-conditioned attention-based edge weights have already been explored in prior literature such as RED-GNN[4], and utilizing query-specific subgraphs is a common practice in the IKGR field [5].

[1] Geng et al., Relational Message Passing for Fully Inductive Knowledge Graph Completion, ICDE 2023.
[2] Lee et al., InGram: Inductive Knowledge Graph Embedding via Relation Graphs, ICML 2023.
[3] Galkin et al., Towards Foundation Models for Knowledge Graph Reasoning, ICLR 2024.
[4] Zhang and Yao, Knowledge Graph Reasoning with Relational Digraph, WWW 2022.
[4] Teru et al., Inductive Relation Prediction by Subgraph Reasoning, ICML 2020.

Questions for authors
  • How is addressing only the entity-inductive setting a more realistic scenario, or what specific advantages does it hold, compared to the setting where both entities and relations can be unseen?

  • What is the core technical novelty of the proposed method beyond the application of invariant learning?

  • In Table 1, while the performance gains on FB15K-237 v2–v4 are relatively marginal or slightly lower than KnowFormer, HiKGR achieves an unusually large jump on FB15K-237 v1 (0.600 vs. 0.466 on MRR). What specific characteristics of the v1 split account for this substantial discrepancy?

  • All in-text citations are missing parentheses around the authors' names, which makes the text difficult to read. Could the authors clarify whether this was intended, or fix the citation formatting in the revision to improve readability?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.