Pith. sign in

REVIEW 4 major objections 6 minor 74 references

Rule Learning for Knowledge Graph Reasoning under Agnostic Distribution Shift

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that reweighting rule-body embeddings to make their dimensions statistically independent lets logical rule learning generalize across shifted query distributions, and supports this with the StableRule framework.

desk verdict A sensible application of stable-learning reweighting to rule learning for KG reasoning, with a useful problem framing but a thin theory section and an OOD evaluation that targets the method's own assumed shift regime. read the letter →

arxiv 2507.05110 v3 pith:VE53QJ6F submitted 2025-07-07 cs.AI

Shixuan Liu , Yue He , Yunfei Wang , Hao Zou , Haoxiang Cheng , Wenjing Yang , Peng Cui , Zhong Liu This is my paper · ORCID
classification cs.AI
keywords knowledgegraphreasoninglogicalrulelearningout-of-distributiongeneralizationdistributionshiftcovariatefeaturedecorrelationsamplereweightinglinkprediction
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Logical rule learning infers missing facts by scoring rules that connect a body of relations to a head relation. The paper argues that when test queries come from shifted distributions, the relation paths seen in training are biased (covariate shift in $P(r_b)$), which distorts rule plausibility scores. It claims that reweighting sampled rule instances so that the latent rule-body embeddings have statistically independent dimensions removes those spurious correlations and restores stable reasoning. The proposed StableRule framework pairs this feature-decorrelation regularizer with an encoder-decoder rule learning network, and experiments on seven benchmark knowledge graphs report higher and more consistent performance across shifted subgroups than twelve comparison methods. If correct, rule-based KG reasoning becomes reliable in settings where user queries drift over time or across populations.

What carries the argument

The load-bearing object is the feature-decorrelation regularizer applied to rule-body embeddings. For a batch of rule instances, the encoder maps each body to a $d$-dimensional vector; the regularizer computes per-sample weights $W$ by minimizing $L_d = \sum_{i=1}^d \sum_{a=1}^{O_d} \sum_{b=1}^{O_d} \| \frac{1}{N} Z^a_{*i}^T \Sigma_W Z^b_{*\neg i} - \frac{1}{N} Z^a_{*i}^T W \cdot \frac{1}{N} Z^b_{*\neg i}^T W \|^2$, which enforces cross-moment independence $E[Z_i^a Z_j^b] = E[Z_i^a]E[Z_j^b]$ between each dimension and the rest. The paper shows the optimal weights factorize the joint density, $w^*_n = \prod_{i=1}^d f(Z_{ni})/f(Z_{n1},\ldots,Z_{nd})$, which decorrelates the reweighted distribution, and then uses the weighted cross-entropy loss to train the encoder-decoder rule scorer. A backtracking-enhanced random walk sampler supplies diverse rule instances, and the encoder uses sliding-window LSTM composition with attention to respect the deductive structure of rule bodies.

What would settle it

Construct a test environment with pure label shift: keep $P(r_b)$ identical to training but flip $P(r_h|r_b)$ for a subset of head entities (e.g., the same path now implies a different relation for footballers than for journalists). If StableRule's accuracy degrades no less than the no-decorrelation ablation under this shift, the claim that decorrelation of body embeddings is sufficient for agnostic shifts is refuted. A second check: compute a non-parametric dependence measure between all dimension pairs of the reweighted embeddings on held-out shifted data; if substantial dependence remains while performance is stable, the paper's stated mechanism is not what is driving the gain.

Watch

Extended reading notes

Core claim

The paper's central claim is that agnostic query shift, whether density shift (uneven subgraph density around queried entities) or pathway shift (changes in compositional reasoning paths), acts on rule learning as a covariate shift in the distribution of rule bodies $P(r_b)$. This biases the plausibility score $P(r_h|r_b)$ that rule learners estimate, because spurious correlations in the body path distribution are absorbed into the learned embeddings. StableRule addresses this by jointly optimizing an encoder $\phi_e$, a decoder $\phi_d$, and sample weights $W$ so that the latent dimensions of body embeddings have minimal statistical dependence under the reweighted distribution; the paper proves that a weight solution exists ($w^*_n = \prod_i f(Z_{ni})/f(Z_{n1},\ldots,Z_{nd})$) and shows empirically that the resulting rules keep higher and less variable accuracy across five shifted test subgroups on DBLP, WN18RR, FB15K-237, YAGO3-10 and smaller benchmarks, including in inductive settings with disjoint entities.

Load-bearing premise

The load-bearing premise is that query shifts act on rule learning mainly as a covariate shift in the distribution of relation paths $P(r_b)$, so removing spurious correlations among rule-body embedding dimensions suffices to restore accurate reasoning; if the shift also changes the mapping from body to head relations, alters relation semantics, or introduces novel relations, the decorrelation mechanism is not targeted at the cause.

Editorial extensions

If this is right

  • Rule-based KG reasoners can be deployed where test queries concentrate on different entity groups than training data, without requiring group labels or recollecting training facts.
  • Mean accuracy across shifted subgroups stays competitive while subgroup variance shrinks, because the reweighting removes correlations that overfit to over-represented paths.
  • In inductive settings where training and test entities are disjoint, the stability gain over other rule-learning methods is larger than in transductive settings.
  • The decorrelation module also reduces selection bias in the path-sampling step itself, so performance remains competitive even when no synthetic shift is injected.
  • The added computation is modest: training time on FB15K-237 is about twice that of NCRL and far below Neural-LP or DRUM, with time complexity $O((T_d\,d\,O_d^2+2)N+d^2)$.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Truncating decorrelation at moment order $O_d$ means the method enforces approximate independence, not exact independence; environments whose spurious correlations live mainly in higher-order moments may need larger $O_d$ or a non-parametric dependence measure to be fully handled.
  • The same reweighting recipe could be applied to other differentiable reasoners with fixed-dimensional intermediate representations, such as path-based GNNs, whenever their embeddings face analogous covariate shift.
  • If a shift changes the semantics of relations themselves or introduces unseen relations, decorrelation of body embeddings addresses the wrong level; the method would need to be combined with relation-adaptation or open-set components.
  • The learned sample weights could serve as a diagnostic: paths with extreme weights in a given test environment point to the specific relations whose sampling is biased, turning the regularizer into an interpretability tool.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper studies logical rule learning for knowledge graph reasoning when test queries are drawn from environments that differ from the training distribution. It defines a new problem formulation, OOD KG reasoning, and argues that query shifts primarily manifest as covariate shift in the relation-path distribution P(rb). The proposed framework, StableRule, combines an encoder-decoder rule learning network with a feature decorrelation regularizer that reweights rule-body embeddings to reduce statistical dependence among embedding dimensions. Experiments are conducted on seven KGs across statistical relation learning, standard KG completion, and inductive link prediction, using both natural subgroup shifts (DBLP country groups) and synthetic test splits constructed by clustering queries according to path-describing vectors. The central claim is that joint optimization of the encoder, decoder, and sample weights yields representations with minimal statistical dependence and thereby improves generalization to shifted queries.

Significance. The problem is timely and the paper appears to be the first to explicitly address distribution shift in logical rule learning for KG reasoning. The method is simple and intuitive, the benchmark construction is thoughtful, and the empirical study is broad, covering transductive and inductive settings, ablations, convergence behavior, and mined-rule inspection. If the mechanism were rigorously supported, the contribution would be a solid advance for the rule-learning branch of KG reasoning. The main weaknesses are that the theoretical result does not connect to the joint optimization actually performed, and the experimental shifts are confined to the covariate-shift regime encoded by path statistics. The paper also overstates its empirical superiority on at least one dataset.

major comments (4)
  1. [§4.1–4.2, Lemma 4.1] The load-bearing claim of Section 4.1 is that joint optimization of φe, φd, and W ensures minimal statistical dependence across embedding dimensions and facilitates generalization to shifted queries. Lemma 4.1 only shows that for a fixed encoder a weight vector exists that decorrelates a kernel estimate of the embedding distribution in the limit; it does not address the alternating optimization in Algorithm 1, where weights are re-fit in an inner loop while φe and φd change in the outer loop, and it gives no bound relating decorrelation to test-time risk under covariate shift in P(rb). The cited stable-learning guarantees in [51,53,59] require conditions that are not verified here, such as linear or fixed feature spaces, known invariant feature sets, or bounded density ratios. The proof also contains notational slips, including lim n→∞ where lim N→∞ is intended and an incomplete denominator in the cross-moment display. This gap is load-bearing because the paper's central claim is about the joint optimization, not merely the existence of a decorrelating reweighting for fixed features.
  2. [§3.3.1 and §5.1] The paper reduces 'agnostic' query shift to covariate shift in P(rb) by assertion rather than derivation. Definition 4 only requires P(eh) to vary across environments, and Section 3.3.1 then argues that such shifts manifest as changes in path-frequency statistics. The synthetic test sets in Section 5.1 are constructed by clustering queries according to path-describing vectors, i.e., along exactly the axis that the decorrelation module corrects. This leaves untested label shift, shifts in relation semantics, and novel-relation settings, and it makes the 'agnostic' claim overbroad. A concrete test would be to build subgroups from head-entity attributes that do not change P(rb), or to hold out entire relations, and to report whether decorrelation still helps.
  3. [§5.3, Table 4; §5.4, Table 6] The results on WN18RR contradict the unqualified 'superior performance' statements. In Table 4, StableRule has MRR 51.01 versus NBFNet 53.87 and A*Net 51.18; in Table 6, its MRR is 48.53 versus NBFNet 52.96 and A*Net 50.88. The text in Section 5.4 says 'our method achieves superior performance' and the abstract claims superior effectiveness, so the claims need to be qualified to rule-based methods or to the datasets where the advantage actually holds.
  4. [§6.4, Table 9] The ablation establishes necessity of the decorrelation term in the tested configurations (a 5.18% average MRR drop for NoDecor), but it does not establish that the specific decorrelation objective is the cause of the gains. No control is presented with generic reweighting, such as inverse propensity weighting on rule-body frequencies or reweighting to balance path lengths and relations. Without such controls, the improvement could plausibly come from the reweighting acting as a regularizer or from the backtracking sampler rather than from the decorrelation mechanism emphasized in the paper. Please add at least one alternative-reweighting baseline.
minor comments (6)
  1. [§4.2, Lemma 4.1] Please fix 'lim n→∞ Ld' to 'lim N→∞ Ld' and correct the denominator in the cross-moment display to the full joint density estimate \hat f(Z_{n1},...,Z_{nd}).
  2. [§5.2, Figure 4] The body-coverage percentages in the upper part of Figure 4 are described in the caption but the figure lacks clear row labels and a legend; please add them so the reader can map each percentage to a rule.
  3. [§6.4, Table 9] Report standard deviations or significance tests for the ablation variants, as Table 4 does for the main results; this would help assess the stability of the 5.18% and 2.99% drops.
  4. [§4.3.2] The negative-sampling rule 'if no relation links eh and ei at any point' is ambiguous; clarify whether a non-closed path of any length triggers a negative sample or only the final step of the walk.
  5. [Appendix C, Eq. (8)] The typesetting of the argmax expression is broken ('argmax X ...'); please reformulate the forward-chaining equation so the summation and argmax are unambiguous.
  6. [Abstract and Section 5.1] The paper says 'seven benchmark KGs' but the DBLP natural-shift experiment uses a single relation; please state this clearly or rephrase the claim.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular step in the derivation chain; the method applies stable-learning decorrelation to rule-body embeddings and validates it on external benchmarks, with only a minor same-group citation for the transfer guarantee.

full rationale

The main derivation (Sections 3.3 and 4) does not reduce to its own inputs. Lemma 4.1 establishes existence of a reweighting W that drives Ld to zero using kernel density estimators and a textbook consistency result [64], and this proof does not assume the OOD generalization claim. The bridge from decorrelation to generalization is taken from stable-learning theory [51,52,53,59]; some of these works share authors with the present paper, but they are prior theoretical results with stated assumptions and are not fitted to the present data, so they count as external support under the review rules. The synthetic subgroups in Section 5.1 are built from path-describing vectors, i.e., statistics of P(rb), which matches the method's assumed shift mechanism; this is an evaluation design choice that may limit external validity, but it does not feed test labels back into training or make the reported improvements true by construction. Sample weights W are optimized on training rule instances only (Algorithm 1) and are not a renamed prediction. The conclusion's stated plan to study additional shift sources confirms the scope is covariate shift in P(rb) rather than a universal OOD guarantee. No equation in the paper equals its own input, and no fitted parameter is presented as a prediction.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The framework depends on two external pillars: the stable-learning theory that decorrelating features improves covariate-shift generalization, and the standard KDE convergence theory used in Lemma 4.1. It introduces no new conceptual entities. The key free parameter is the moment order O_d, set to 2 by hand; the maximum rule length l=3 is a design constraint. The central claim does not depend on a single fitted constant, but the empirical results are sensitive to the chosen hyperparameters.

free parameters (2)
  • moment order O_d = 2
    Controls the order of moments in the decorrelation loss (Eq. 5). Set to 2 by hand; Section 4.2 states it is treated as a hyperparameter.
  • maximum rule length l = 3
    Limits the length of rule bodies learned; set to 3 in all experiments (Appendix E). Hand-chosen constraint.
assumptions (3)
  • domain assumption Test-time query shifts primarily induce covariate shift in the relation path distribution P(rb)
    Section 3.3.1 argues this reduction; the whole decorrelation approach depends on it.
  • domain assumption Decorrelating learned representation dimensions via sample reweighting improves OOD generalization for nonlinear LSTM and attention encoders
    Adopted from stable learning theory [51,18,53]; the paper cites this as established for images and graphs but does not prove it for rule body embeddings.
  • standard math Kernel density estimators are consistent
    Used in Lemma 4.1 to show existence of reweighting weights; standard result from [64].

how reviews work

0 comments
Cite this review

Pith. "Pith review of Rule Learning for Knowledge Graph Reasoning under Agnostic Distribution Shift." pith.science (2026). https://pith.science/paper/VE53QJ6F

@misc{pith2026250705110,
  author       = {Pith},
  title        = {Pith review of: Rule Learning for Knowledge Graph Reasoning under Agnostic Distribution Shift},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VE53QJ6F}},
  note         = {Machine review of arXiv:2507.05110}
}
read the original abstract

Logical rule learning, a prominent category of knowledge graph (KG) reasoning methods, constitutes a critical research area aimed at learning explicit rules from observed facts to infer missing knowledge. However, like all KG reasoning methods, rule learning suffers from a critical weakness-its dependence on the I.I.D. assumption. This assumption can easily be violated due to selection bias during training or agnostic distribution shifts during testing (e.g., as in query shift scenarios), ultimately undermining model performance and reliability. To enable robust KG reasoning in wild environments, this study investigates logical rule learning in the presence of agnostic test-time distribution shifts. We formally define this challenge as out-of-distribution (OOD) KG reasoning-a previously underexplored problem, and propose the Stable Rule Learning (StableRule) framework as a solution. StableRule is an end-to-end framework that combines feature decorrelation with rule learning network, to enhance OOD generalization in KG reasoning. By leveraging feature decorrelation, StableRule mitigates the adverse effects of covariate shifts arising in OOD scenarios, improving the robustness of the rule learning network. Extensive experiments on seven benchmark KGs demonstrate the framework's superior effectiveness and stability across diverse heterogeneous environments, highlighting its practical significance for real-world applications.

Figures

Figures reproduced from arXiv: 2507.05110 by the authors.

Figure 1
Figure 1. Illustration of KG Reasoning under Query Shift. Solid edges denote observed facts in the training graph, while dashed edges represent [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of StableRule. We initially generate a batch of rule instances [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Visualization of the partitioning of test set over one relation. We can see the queries are divided into distinct groups. [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Results on DBLP over subgroups. The above part illustrates the variations in body coverage (which gauges the frequency of each body, [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Loss curves on six datasets. Each curve is averaged over 3 runs. [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Comparison of feature correlation matrices before (left) [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

74 extracted references · 65 canonical work pages

  1. [1]

    Dbpedia: A nucleus for a web of open data,

    S. Auer, C. Bizer, G. Kobilarov, J. Lehmann, R. Cyganiak, and Z. Ives, “Dbpedia: A nucleus for a web of open data,” in The semantic web. Springer, 2007, pp. 722–735

    work page 2007

  2. [2]

    Yago: a core of semantic knowledge,

    F. M. Suchanek, G. Kasneci, and G. Weikum, “Yago: a core of semantic knowledge,” in Proceedings of the 16th international con- ference on World Wide Web, 2007, pp. 697–706

    work page 2007

  3. [3]

    Hkgb: an inclusive, extensible, intelligent, semi- auto-constructed knowledge graph framework for healthcare with clinicians’ expertise incorporated,

    Y. Zhang, M. Sheng, R. Zhou, Y. Wang, G. Han, H. Zhang, C. Xing, and J. Dong, “Hkgb: an inclusive, extensible, intelligent, semi- auto-constructed knowledge graph framework for healthcare with clinicians’ expertise incorporated,” Information Processing & Man- agement, vol. 57, no. 6, p. 102324, 2020

    work page 2020

  4. [4]

    Kgat: Knowledge graph attention network for recommendation,

    X. Wang, X. He, Y. Cao, M. Liu, and T.-S. Chua, “Kgat: Knowledge graph attention network for recommendation,” inProceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, 2019, pp. 950–958

    work page 2019

  5. [5]

    Relational learning analysis of social politics using knowledge graph embedding,

    B. Abu-Salih, M. Al-Tawil, I. Aljarah, H. Faris, P . Wongthongtham, K. Y. Chan, and A. Beheshti, “Relational learning analysis of social politics using knowledge graph embedding,” Data Mining and Knowledge Discovery, vol. 35, no. 4, pp. 1497–1536, 2021

    work page 2021

  6. [6]

    A comprehensive overview of knowledge graph completion,

    T. Shen, F. Zhang, and J. Cheng, “A comprehensive overview of knowledge graph completion,” Knowledge-Based Systems , p. 109597, 2022

    work page 2022

  7. [7]

    A review: Knowledge reasoning over knowledge graph,

    X. Chen, S. Jia, and Y. Xiang, “A review: Knowledge reasoning over knowledge graph,” Expert Systems with Applications, vol. 141, p. 112948, 2020

    work page 2020

  8. [8]

    A survey on knowl- edge graph embedding: Approaches, applications and bench- marks,

    Y. Dai, S. Wang, N. N. Xiong, and W. Guo, “A survey on knowl- edge graph embedding: Approaches, applications and bench- marks,” Electronics, vol. 9, no. 5, p. 750, 2020

    work page 2020

Show all 74 references
  1. [9]

    Differentiable learning of logical rules for knowledge base reasoning,

    F. Yang, Z. Yang, and W. W. Cohen, “Differentiable learning of logical rules for knowledge base reasoning,” Advances in neural information processing systems, vol. 30, 2017

    2017

  2. [10]

    Sparsity and noise: Where knowledge graph embeddings fall short,

    J. Pujara, E. Augustine, and L. Getoor, “Sparsity and noise: Where knowledge graph embeddings fall short,” in Proceedings of the 2017 conference on empirical methods in natural language processing , 2017, pp. 1751–1756

    2017

  3. [11]

    From local structures to size generalization in graph neural networks,

    G. Yehudai, E. Fetaya, E. Meirom, G. Chechik, and H. Maron, “From local structures to size generalization in graph neural networks,” in International Conference on Machine Learning. PMLR, 2021, pp. 11 975–11 986

    2021

  4. [12]

    Size-invariant graph representations for graph classification extrapolations,

    B. Bevilacqua, Y. Zhou, and B. Ribeiro, “Size-invariant graph representations for graph classification extrapolations,” in Interna- tional Conference on Machine Learning. PMLR, 2021, pp. 837–851

    2021

  5. [13]

    Logical rule learning,

    K. Cheng and Y. Sun, “Logical rule learning,” Knowledge Graph Reasoning: A Neuro-Symbolic Perspective, pp. 107–147, 2024

    2024

  6. [14]

    Ood-gnn: Out-of- distribution generalized graph neural network,

    H. Li, X. Wang, Z. Zhang, and W. Zhu, “Ood-gnn: Out-of- distribution generalized graph neural network,” IEEE Transactions on Knowledge and Data Engineering, 2022

    2022

  7. [15]

    Rnnlogic: Learning logic rules for reasoning on knowledge graphs,

    M. Qu, J. Chen, L.-P . Xhonneux, Y. Bengio, and J. Tang, “Rnnlogic: Learning logic rules for reasoning on knowledge graphs,” in International Conference on Learning Representations, 2020

    2020

  8. [16]

    Offline reinforcement learning: Tutorial, review, and perspectives on open problems,

    S. Levine, A. Kumar, G. Tucker, and J. Fu, “Offline reinforcement learning: Tutorial, review, and perspectives on open problems,” arXiv preprint arXiv:2005.01643, 2020

    2005 arXiv

  9. [17]

    Pretrained transformers improve out-of-distribution robustness,

    D. Hendrycks, X. Liu, E. Wallace, A. Dziedzic, R. Krishnan, and D. Song, “Pretrained transformers improve out-of-distribution robustness,” arXiv preprint arXiv:2004.06100, 2020

    2004 arXiv

  10. [18]

    Deep stable learning for out-of-distribution generalization,

    X. Zhang, P . Cui, R. Xu, L. Zhou, Y. He, and Z. Shen, “Deep stable learning for out-of-distribution generalization,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 5372–5382

    2021

  11. [19]

    Rlogic: Recursive logical rule learning from knowledge graphs,

    K. Cheng, J. Liu, W. Wang, and Y. Sun, “Rlogic: Recursive logical rule learning from knowledge graphs,” in Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , 2022, pp. 179–189

    2022

  12. [20]

    The mean and variance of the distribution of shortest path lengths of random regular graphs,

    I. Tishby, O. Biham, R. K ¨uhn, and E. Katzav, “The mean and variance of the distribution of shortest path lengths of random regular graphs,” Journal of Physics A: Mathematical and Theoretical , vol. 55, no. 26, p. 265005, 2022

    2022

  13. [21]

    Translating embeddings for modeling multi- relational data,

    A. Bordes, N. Usunier, A. Garcia-Duran, J. Weston, and O. Yakhnenko, “Translating embeddings for modeling multi- relational data,” Advances in neural information processing systems , vol. 26, 2013

    2013

  14. [22]

    Knowledge graph embedding by translating on hyperplanes,

    Z. Wang, J. Zhang, J. Feng, and Z. Chen, “Knowledge graph embedding by translating on hyperplanes,” in Proceedings of the AAAI conference on artificial intelligence, vol. 28, no. 1, 2014

    2014

  15. [23]

    Learning entity and relation embeddings for knowledge graph completion,

    Y. Lin, Z. Liu, M. Sun, Y. Liu, and X. Zhu, “Learning entity and relation embeddings for knowledge graph completion,” in Proceedings of the AAAI conference on artificial intelligence , vol. 29, no. 1, 2015

    2015

  16. [24]

    Rotate: Knowledge graph embedding by relational rotation in complex space,

    Z. Sun, Z.-H. Deng, J.-Y. Nie, and J. Tang, “Rotate: Knowledge graph embedding by relational rotation in complex space,” arXiv preprint arXiv:1902.10197, 2019

    1902 arXiv

  17. [25]

    A three-way model for collective learning on multi-relational data

    M. Nickel, V . Tresp, H.-P . Kriegel et al. , “A three-way model for collective learning on multi-relational data.” in Icml, vol. 11, no. 10.5555, 2011, pp. 3 104 482–3 104 584

    2011

  18. [26]

    Embedding entities and relations for learning and inference in knowledge bases,

    B. Yang, W.-t. Yih, X. He, J. Gao, and L. Deng, “Embedding entities and relations for learning and inference in knowledge bases,” arXiv preprint arXiv:1412.6575, 2014

    2014 arXiv

  19. [27]

    Complex embeddings for simple link prediction,

    T. Trouillon, J. Welbl, S. Riedel, ´E. Gaussier, and G. Bouchard, “Complex embeddings for simple link prediction,” in International conference on machine learning. PMLR, 2016, pp. 2071–2080

    2016

  20. [28]

    Modeling relational data with graph convolu- tional networks,

    M. Schlichtkrull, T. N. Kipf, P . Bloem, R. Van Den Berg, I. Titov, and M. Welling, “Modeling relational data with graph convolu- tional networks,” in The semantic web: 15th international conference, ESWC 2018, Heraklion, Crete, Greece, June 3–7, 2018, proceedings 15 . Springe...

    2018

  21. [29]

    Robust embedding with multi-level structures for link prediction

    Z. Wang, Z. Ren, C. He, P . Zhang, and Y. Hu, “Robust embedding with multi-level structures for link prediction.” in IJCAI, 2019, pp. 5240–5246

    2019

  22. [30]

    Composition- based multi-relational graph convolutional networks,

    S. Vashishth, S. Sanyal, V . Nitin, and P . Talukdar, “Composition- based multi-relational graph convolutional networks,” in Interna- tional Conference on Learning Representations

  23. [31]

    Inductive relation prediction by subgraph reasoning,

    K. Teru, E. Denis, and W. Hamilton, “Inductive relation prediction by subgraph reasoning,” in International Conference on Machine Learning. PMLR, 2020, pp. 9448–9457. IEEE T -PAMI SUBMISSION 15

    2020

  24. [32]

    Neural bellman- ford networks: A general graph neural network framework for link prediction,

    Z. Zhu, Z. Zhang, L.-P . Xhonneux, and J. Tang, “Neural bellman- ford networks: A general graph neural network framework for link prediction,” Advances in neural information processing systems , vol. 34, pp. 29 476–29 490, 2021

    2021

  25. [33]

    A* net: A scalable path-based reasoning approach for knowledge graphs,

    Z. Zhu, X. Yuan, M. Galkin, L.-P . Xhonneux, M. Zhang, M. Gazeau, and J. Tang, “A* net: A scalable path-based reasoning approach for knowledge graphs,” Advances in Neural Information Processing Systems, vol. 36, pp. 59 323–59 336, 2023

    2023

  26. [34]

    Knowledge graph reasoning with relational digraph,

    Y. Zhang and Q. Yao, “Knowledge graph reasoning with relational digraph,” in Proceedings of the ACM web conference 2022 , 2022, pp. 912–924

    2022

  27. [35]

    Inductive logic programming: Theory and methods,

    S. Muggleton and L. De Raedt, “Inductive logic programming: Theory and methods,” The Journal of Logic Programming , vol. 19, pp. 629–679, 1994

    1994

  28. [36]

    S. H. Muggleton, C. Feng et al., Efficient induction of logic programs . Turing Institute, 1990

    1990

  29. [37]

    End-to-end differentiable proving,

    T. Rockt ¨aschel and S. Riedel, “End-to-end differentiable proving,” Advances in neural information processing systems , vol. 30, 2017

    2017

  30. [38]

    Drum: End-to-end differentiable rule mining on knowledge graphs,

    A. Sadeghian, M. Armandpour, P . Ding, and D. Z. Wang, “Drum: End-to-end differentiable rule mining on knowledge graphs,” Ad- vances in Neural Information Processing Systems , vol. 32, pp. 15 347– 15 357, 2019

    2019

  31. [39]

    Neural compositional rule learning for knowledge graph reasoning,

    K. Cheng, N. Ahmed, and Y. Sun, “Neural compositional rule learning for knowledge graph reasoning,” in The Eleventh Inter- national Conference on Learning Representations, 2022

    2022

  32. [40]

    Learning to gener- alize: Meta-learning for domain generalization,

    D. Li, Y. Yang, Y.-Z. Song, and T. Hospedales, “Learning to gener- alize: Meta-learning for domain generalization,” in Proceedings of the AAAI conference on artificial intelligence , vol. 32, no. 1, 2018

    2018

  33. [41]

    Domain generalization with adversarial feature learning,

    H. Li, S. J. Pan, S. Wang, and A. C. Kot, “Domain generalization with adversarial feature learning,” in Proceedings of the IEEE confer- ence on computer vision and pattern recognition , 2018, pp. 5400–5409

    2018

  34. [42]

    Domain generalization via model-agnostic learning of semantic features,

    Q. Dou, D. Coelho de Castro, K. Kamnitsas, and B. Glocker, “Domain generalization via model-agnostic learning of semantic features,” Advances in neural information processing systems , vol. 32, 2019

    2019

  35. [43]

    Domain generalization via multidomain discriminant analysis,

    S. Hu, K. Zhang, Z. Chen, and L. Chan, “Domain generalization via multidomain discriminant analysis,” in Uncertainty in Artificial Intelligence. PMLR, 2020, pp. 292–302

    2020

  36. [44]

    Efficient domain gener- alization via common-specific low-rank decomposition,

    V . Piratla, P . Netrapalli, and S. Sarawagi, “Efficient domain gener- alization via common-specific low-rank decomposition,” in Inter- national Conference on Machine Learning . PMLR, 2020, pp. 7728– 7738

    2020

  37. [45]

    Learning to op- timize domain specific normalization for domain generalization,

    S. Seo, Y. Suh, D. Kim, G. Kim, J. Han, and B. Han, “Learning to op- timize domain specific normalization for domain generalization,” in Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXII 16. Springer, 2020, pp. 68–83

    2020

  38. [46]

    Domain generalization by solving jigsaw puzzles,

    F. M. Carlucci, A. D’Innocente, S. Bucci, B. Caputo, and T. Tom- masi, “Domain generalization by solving jigsaw puzzles,” in Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019, pp. 2229–2238

    2019

  39. [47]

    Generalizing across domains via cross-gradient training,

    S. Shankar, V . Piratla, S. Chakrabarti, S. Chaudhuri, P . Jyothi, and S. Sarawagi, “Generalizing across domains via cross-gradient training,” arXiv preprint arXiv:1804.10745, 2018

    2018 arXiv

  40. [48]

    Generalizing to unseen domains via adversarial data augmentation,

    R. Volpi, H. Namkoong, O. Sener, J. C. Duchi, V . Murino, and S. Savarese, “Generalizing to unseen domains via adversarial data augmentation,” Advances in neural information processing systems , vol. 31, 2018

    2018

  41. [49]

    Episodic training for domain generalization,

    D. Li, J. Zhang, Y. Yang, C. Liu, Y.-Z. Song, and T. M. Hospedales, “Episodic training for domain generalization,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019, pp. 1446–1455

    2019

  42. [50]

    Invariant risk minimization,

    M. Arjovsky, L. Bottou, I. Gulrajani, and D. Lopez-Paz, “Invariant risk minimization,” arXiv preprint arXiv:1907.02893, 2019

    1907 arXiv

  43. [51]

    Stable prediction with model misspecification and agnostic distribution shift,

    K. Kuang, R. Xiong, P . Cui, S. Athey, and B. Li, “Stable prediction with model misspecification and agnostic distribution shift,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 34, no. 04, 2020, pp. 4485–4492

    2020

  44. [52]

    Stable learning via sample reweighting,

    Z. Shen, P . Cui, T. Zhang, and K. Kunag, “Stable learning via sample reweighting,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 34, no. 04, 2020, pp. 5692–5699

    2020

  45. [53]

    Why stable learning works? a theory of covariate shift generalization,

    R. Xu, P . Cui, Z. Shen, X. Zhang, and T. Zhang, “Why stable learning works? a theory of covariate shift generalization,” arXiv preprint arXiv:2111.02355, vol. 2, 2021

    2021 arXiv

  46. [54]

    Decorrelate irrelevant, purify relevant: Overcome textual spurious correlations from a feature perspective,

    S. Dou, R. Zheng, T. Wu, S. Gao, J. Shan, Q. Zhang, Y. Wu, and X. Huang, “Decorrelate irrelevant, purify relevant: Overcome textual spurious correlations from a feature perspective,” arXiv preprint arXiv:2202.08048, 2022

    2022 arXiv

  47. [55]

    Sound and complete forward and backward chainings of graph rules,

    E. Salvat and M.-L. Mugnier, “Sound and complete forward and backward chainings of graph rules,” in International Conference on Conceptual Structures. Springer, 1996, pp. 248–262

    1996

  48. [56]

    Amie: association rule mining under incomplete evidence in ontological knowledge bases,

    L. A. Gal ´arraga, C. Teflioudi, K. Hose, and F. Suchanek, “Amie: association rule mining under incomplete evidence in ontological knowledge bases,” in Proceedings of the 22nd international conference on World Wide Web, 2013, pp. 413–422

    2013

  49. [57]

    Robust covariate shift regression,

    X. Chen, M. Monfort, A. Liu, and B. D. Ziebart, “Robust covariate shift regression,” in Artificial Intelligence and Statistics . PMLR, 2016, pp. 1270–1279

    2016

  50. [58]

    Learning under nonstationarity: covariate shift and class-balance change,

    M. Sugiyama, M. Yamada, and M. C. du Plessis, “Learning under nonstationarity: covariate shift and class-balance change,” Wiley Interdisciplinary Reviews: Computational Statistics , vol. 5, no. 6, pp. 465–477, 2013

    2013

  51. [59]

    A theoretical anal- ysis on independence-driven importance weighting for covariate- shift generalization,

    R. Xu, X. Zhang, Z. Shen, T. Zhang, and P . Cui, “A theoretical anal- ysis on independence-driven importance weighting for covariate- shift generalization,” in International Conference on Machine Learn- ing. PMLR, 2022, pp. 24 803–24 829

    2022

  52. [60]

    An empirical study on robustness to spurious correlations using pre-trained language models,

    L. Tu, G. Lalwani, S. Gella, and H. He, “An empirical study on robustness to spurious correlations using pre-trained language models,” Transactions of the Association for Computational Linguistics, vol. 8, pp. 621–633, 2020

    2020

  53. [61]

    When does e (xk· yl)= e (xk)· e (yl) imply independence?

    T. M. Bisgaard and Z. Sasv ´ari, “When does e (xk· yl)= e (xk)· e (yl) imply independence?” Statistics & probability letters, vol. 76, no. 11, pp. 1111–1116, 2006

    2006

  54. [62]

    Approximate residual balancing: debiased inference of average treatment effects in high dimensions,

    S. Athey, G. W. Imbens, and S. Wager, “Approximate residual balancing: debiased inference of average treatment effects in high dimensions,” Journal of the Royal Statistical Society Series B: Statisti- cal Methodology, vol. 80, no. 4, pp. 597–623, 2018

    2018

  55. [63]

    Covariate balancing propensity score for a continuous treatment: Application to the efficacy of political advertisements,

    C. Fong, C. Hazlett, and K. Imai, “Covariate balancing propensity score for a continuous treatment: Application to the efficacy of political advertisements,” The Annals of Applied Statistics , vol. 12, no. 1, pp. 156–177, 2018

    2018

  56. [64]

    Hollander, D

    M. Hollander, D. A. Wolfe, and E. Chicken, Nonparametric statistical methods. John Wiley & Sons, 2013

    2013

  57. [65]

    Spitzer, Principles of random walk

    F. Spitzer, Principles of random walk . Springer Science & Business Media, 2013, vol. 34

    2013

  58. [66]

    Learning distributed representations of con- cepts,

    G. E. Hinton et al., “Learning distributed representations of con- cepts,” in Proceedings of the eighth annual conference of the cognitive science society, vol. 1. Amherst, MA, 1986, p. 12

    1986

  59. [67]

    Statistical predicate invention,

    S. Kok and P . Domingos, “Statistical predicate invention,” in Proceedings of the 24th international conference on Machine learning , 2007, pp. 433–440

    2007

  60. [68]

    Convo- lutional 2d knowledge graph embeddings,

    T. Dettmers, P . Minervini, P . Stenetorp, and S. Riedel, “Convo- lutional 2d knowledge graph embeddings,” in Proceedings of the AAAI conference on artificial intelligence, vol. 32, no. 1, 2018

    2018

  61. [69]

    Observed versus latent features for knowledge base and text inference,

    K. Toutanova and D. Chen, “Observed versus latent features for knowledge base and text inference,” in Proceedings of the 3rd workshop on continuous vector space models and their compositionality , 2015, pp. 57–66

    2015

  62. [70]

    Arnetminer: extraction and mining of academic social networks,

    J. Tang, J. Zhang, L. Yao, J. Li, L. Zhang, and Z. Su, “Arnetminer: extraction and mining of academic social networks,” inProceedings of the 14th ACM SIGKDD international conference on Knowledge discovery and data mining, 2008, pp. 990–998

    2008

  63. [71]

    Geotext: an intelligent dynamic geometry textbook,

    X. Chen, D. Wang, and T. Zhao, “Geotext: an intelligent dynamic geometry textbook,” ACM Communications in Computer Algebra , vol. 46, no. 3/4, pp. 171–175, 2013

    2013

  64. [72]

    Pytorch: An im- perative style, high-performance deep learning library,

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga et al., “Pytorch: An im- perative style, high-performance deep learning library,” Advances in neural information processing systems, vol. 32, 2019

    2019

  65. [73]

    Adam: A method for stochastic optimiza- tion,

    D. P . Kingma and J. Ba, “Adam: A method for stochastic optimiza- tion,” arXiv preprint arXiv:1412.6980, 2014

    2014 arXiv

  66. [74]

    Heterogeneous information networks: the past, the present, and the future,

    Y. Sun, J. Han, X. Yan, P . S. Yu, and T. Wu, “Heterogeneous information networks: the past, the present, and the future,” Proceedings of the VLDB Endowment , vol. 15, no. 12, pp. 3807–3811, 2022. IEEE T -PAMI SUBMISSION 16 Shixuan Liu received his B.S. and Ph.D. de- grees fro...

    2022

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.