Knowledge Graph Completion by Intermediate Variables Regularization

Changyi Xiao; Yixin Cao

REVIEW 3 major objections 4 minor 29 references

Knowledge Graph Completion by Intermediate Variables Regularization

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper proposes a regularizer that bounds tensor rank for knowledge graph completion models and shows empirical gains.

desk verdict New regularizer with real empirical gains, but the proof of the trace-norm bound is false for α>2, and that's the regime used in the main experiments. read the letter →

arxiv 2506.02749 v1 pith:ST2MFSZ5 submitted 2025-06-03 cs.LG cs.AI

Changyi Xiao , Yixin Cao This is my paper

classification cs.LGcs.AI MSC 15A6968T30

keywords knowledgegraphcompletiontensordecompositionregularizationtracenormoverfitting

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

The paper proposes Intermediate Variables Regularization (IVR), a penalty on the norms of intermediate quantities that arise when a tensor-decomposition-based knowledge graph completion model computes its predicted score tensor. The authors claim that this penalty upper-bounds the overlapped trace norm of the predicted tensor, a standard surrogate for tensor rank, and therefore directly targets overfitting by encouraging correlated entity and relation embeddings. They verify the claim empirically for six TDB models (CP, ComplEx, SimplE, ANALOGY, QuatE, TuckER) on three standard KGC datasets, reporting MRR improvements over existing regularizers where those regularizers are applicable.

What carries the argument

The key object is the overlapped trace norm L(X;α) = ‖X(1)‖_∗^{α/2} + ‖X(2)‖_∗^{α/2} + ‖X(3)‖_∗^{α/2}, the sum of trace norms of the three mode-unfoldings of the predicted tensor, which is used as a convex surrogate for tensor rank. The proof machinery is Lemma 1, an identity for the α-th power of the trace norm of a matrix Z as a minimum, over factorizations Z = UV^T, of a balanced combination of the 2α-th powers of the Frobenius norms of U and V. Applying that identity to each mode-unfolding of X turns the overlapped trace norm into sums of Frobenius norms of exactly the intermediate variables that IVR penalizes, which is what lets the paper claim that IVR is an upper bound on L(X;α).

What would settle it

A concrete counterexample with X ∈ $R^{{1×1×1}}$, the value X = 2, and a CP decomposition with two components (D = 2) at α = 4 gives a left-hand side of 24 against a right-hand side of 12 in the key inequality, contradicting Proposition 1; equally, computing L(X;α) and the IVR penalty for a randomly generated tensor at α = 3.0 and checking whether the claimed inequality holds would settle the matter.

Watch

Extended reading notes

Core claim

The central claim is that minimizing the norms of intermediate variables in the many ways of computing the tensor decomposition output — the embeddings themselves, their pairwise Kronecker products, the core tensor applied to one mode, and the core tensor applied to two modes — provides a tractable uniform regularizer for TDB models. The theoretical support is Proposition 1: for any decomposition of the predicted tensor, the overlapped trace norm L(X;α) is upper-bounded, up to a constant, by the IVR terms that pair embedding norms with norms of the two-mode products W×2R×3T, W×3T×1H, W×1H×2R. The regularizer also includes the complementary pair of terms (products of embedding norms with one-mode products), giving the full IVR objective. Empirically, IVR improves mean reciprocal rank over the F2, N3, and DURA baselines for the tested models, with the largest gains on models (SimplE, ANALOGY, QuatE, TuckER) that previously had no strong specialized regularizer.

Load-bearing premise

The theoretical guarantee depends on the inequality that the α/2-th power of a trace norm of a sum of matrices is bounded by the sum of the α/2-th powers of the summands, which fails when α > 2 and the bounds used in the main experiments with α > 2 are therefore not covered by the proof as written.

Editorial extensions

If this is right

IVR gives the first uniform regularizer that applies to most tensor-decomposition KGC models, not only CP-derived models like N3 and DURA.
If the theoretical bound holds, minimizing IVR minimizes a rank surrogate of the predicted tensor, yielding a principled and not merely heuristic form of overfitting control.
The ablation shows that the bound in Proposition 1 (IVR-1) alone nearly matches the full IVR, so users can trade a single regularization coefficient against slightly lower performance in cases where hyperparameter budget is tight.
The framework suggests new TDB models could be designed by choosing the partition P and core tensor W while inheriting IVR as a ready-made regularizer.

Reading between the lines

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

A testable extension: compute L(X;α) directly for a trained model and compare the value against the IVR penalty; the paper does this only on the small Kinship dataset, and doing it on WN18RR with randomized subsampling of unfoldings could check the bound's practical tightness at scale.
The regularizer's correlation narrative suggests it should most help datasets with many symmetric or inverse relation pairs and well-clustered entities; an experiment that correlates the MRR gain of IVR with the prevalence of such relation patterns across datasets could sharpen when to use it.
If Proposition 1 fails for α > 2 as a counterexample indicates, the practical gains from IVR may still be real but would need a different theoretical justification, for instance that the penalty favors small norms rather than strictly small rank.
The general-form unification in Eq. (2) could be used to instantiate new models with custom core tensors, with IVR as a ready-made regularizer.

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.

Desk Editor's Note

read the letter

You should know two things about arXiv:2506.02749. First, the empirical work is real: a new regularizer (IVR) applied to six tensor-decomposition KGC models, with consistent MRR gains over F2, N3, and DURA on WN18RR, FB15k-237, and YAGO3-10, plus code and ablations. Second, the theoretical justification that anchors the paper is broken.

What's genuinely new: they unify CP, ComplEx, SimplE, ANALOGY, QuatE, TuckER under Eq.(2), a block-term decomposition with shared core, and propose regularizing norms of intermediate variables. That's a reasonable idea, and the experimental implementation is careful. The gains over DURA are modest but real for several models, and the ablations (IVR-1 vs IVR-2, the Kinship trace-norm measurements) are a good attempt to check the mechanism.

The problem: Proposition 1 claims 2√(λ1λ4) L(X;α) ≤ RHS of Eq.(5). The proof in Appendix B bounds ||Σ_d A_d||_*^{α/2} by Σ_d ||A_d||_*^{α/2}. That step is valid only for α/2 ≤ 1. For α > 2, the inequality goes the other way. The stress-test counterexample is correct: a 1×1×1 tensor X=2, D=2, α=4 gives 24 ≤ 12. So the proposition is false as stated, not merely underproved. Table 6 runs every model except TuckER at α = 3.0 or 2.25, so the theory does not cover the main experiments. Proposition 2 has the same step. The Kinship check uses TuckER at α=2, which is the regime where the proof would go through, so it doesn't validate the rest.

This is a load-bearing flaw: the paper's central claim is that IVR promotes low trace norm, and that claim is unsupported for the hyperparameters they actually use. The regularizer might still work well empirically, and I wouldn't be surprised if a corrected analysis (say, with an extra dependence on D/P, or restricted to α=2) can be made. But as written, the theory is not just loose; it's contradicted.

Who is this for? Someone working on KGC regularization would want to know about the empirical results and the general form. But I would not rely on the theoretical assertions. A serious referee should see this paper, but it needs a substantial revision before acceptance.

Referee Report

3 major / 4 minor

Summary. The paper proposes Intermediate Variables Regularization (IVR), a regularizer for tensor-decomposition-based (TDB) knowledge graph completion models. The authors unify CP, ComplEx, SimplE, ANALOGY, QuatE, and TuckER under a general block-term form (Eq. 2), and define IVR as a weighted sum of Frobenius-norm powers over the intermediate variables that arise in different ways of computing the predicted tensor. The central theoretical claim is that IVR upper-bounds the overlapped trace norm L(X; α) (Propositions 1 and 2), thereby promoting low tensor rank and reducing overfitting. Experiments on WN18RR, FB15k-237, and YAGO3-10 show consistent MRR improvements over F2, N3, and DURA baselines, and a Kinship experiment measures reductions in L(X; α).

Significance. If the theoretical claim were sound, IVR would be a valuable and broadly applicable regularizer, since it covers several model families and outperforms existing regularizers in Table 1. The paper also contributes a unified view of TDB models (Eq. 2) and a logical-rule analysis (Appendix A), which are useful organizational results. The code release and the consistent empirical gains are strengths. However, the main advertised contribution is the theoretical justification, and the proof of the central upper bound is invalid in the experimental regime that the paper actually uses. The empirical results alone, while suggestive, do not compensate for a false theorem at the core of the paper.

major comments (3)

[Appendix B, proof of Proposition 1 (Eq. 7)] The proof uses the step 2L(X;α) ≤ Σ_d ||H:d:(W(1)(T:d:⊗R:d:)^T)||_*^{α/2} + ... , which relies on the inequality (Σ_d a_d)^p ≤ Σ_d a_d^p for p = α/2. This inequality holds only for p ≤ 1; for p > 1 the direction is reversed. The main experiments use α = 3.0 or 2.25 for every model except TuckER (Table 6), so p > 1 throughout the reported experimental regime. The proposition is false as stated: for the 1×1×1 tensor X = 2 with a CP decomposition (W = 1, D = 2, H:d: = R:d: = T:d: = [1] for d = 1,2), α = 4, and λ1 = λ4 = 1, the left side of Eq. (5) is 2√(λ1λ4)L(X;4) = 2(4+4+4) = 24, while the right side is 6 + 6 = 12, giving 24 ≤ 12. The same power inequality appears in the proof of Proposition 2 (after Eq. (8)), so both upper-bound results are unsupported for α > 2.
[Appendix B, tightness construction in Proposition 1] The claimed equality construction is dimensionally inconsistent with the model class in Eq. (2). The proof sets H:1: = (λ1/λ4)^{-1/α} U1√Σ1 with U1 ∈ R^{n1×r1}, and analogously R:1:, T:1:, and W ∈ R^{r1×r2×r3}. In the general form Eq. (2), however, the factors must satisfy H:d: ∈ R^{n1×P}, R:d: ∈ R^{n2×P}, T:d: ∈ R^{n3×P}, and W ∈ R^{P×P×P}, where P is the number of parts. The proof states 'Let P = D', but the n-ranks r1, r2, r3 are determined by the tensor X and are not equal to the embedding dimension D in general. Thus the proposed decomposition does not belong to the model class, and the claimed tightness of Eq. (5) is not established.
[Section 4.4 and Table 3] The empirical verification of the upper bound on Kinship uses TuckER only, and Table 6 reports α = 2.0 for TuckER. At α = 2 the power inequality is an equality, so this experiment tests the theorem only in the regime where it is valid. It does not provide any evidence for the α > 2 settings used for CP, ComplEx, SimplE, ANALOGY, and QuatE in Table 1, where the theorem is false. The sentence in Section 4.4 that the results 'confirm the reliability of our theoretical analysis' is therefore not supported for the main experimental configuration.

minor comments (4)

[Section 3.3] In the paragraph after Proposition 2, the text says 'the r.h.s. of Eq.(4) is an upper bound'; the reference should be to Eq.(6), since Eq.(4) is the weighted regularizer, not the bound in Proposition 2.
[Table 4 and Section 4.1] The dataset name is written inconsistently as 'YGAO3-10' in Table 1 and Table 4 and as 'YAGO3-10' in the text of Section 4.1; please unify the spelling.
[Appendix B, Lemma 1] In the proof of Lemma 1, the expression '∥Z∥α∗ = Tr(Σ)α' should be parenthesized as '(Tr(Σ))^α' to avoid confusion with Tr(Σ^α).
[Section 3.2] The notation H:d: is used for the d-th slice of the reshaped embedding, but in Eq. (4) the slices are written as Hid:, Rjd:, Tkd: for a single triplet; the relationship between these notations could be stated explicitly for readability.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the IVR upper-bound theorem is derived from an external matrix-factorization lemma, and the experimental check measures the target trace norm independently.

full rationale

The paper's central derivation is Proposition 1 (Eq. 5), which claims that the IVR regularizer (Eq. 3) upper-bounds the overlapped trace norm L(X; alpha). The proof in Appendix B reduces this to Lemma 1, credited to Ciliberto et al. (2017) and re-proved in the appendix using SVD and norm inequalities. This lemma is an external, parameter-free factorization result; it is not defined in terms of L(X; alpha) or the IVR regularizer, and the proposition does not fit any parameter to the trace norm. The equality half of Proposition 1 is an explicit existence construction using compact SVDs, not a fitted input disguised as a prediction. The Kinship experiments (Table 3) compute L(X; alpha) directly from the trained predicted tensor, and this quantity is not part of the training loss; it is a post-hoc measurement, so the result is not statistically forced by construction. The paper contains no load-bearing self-citations: the proof relies on Ciliberto et al. (2017) and standard tensor literature, not on the present authors' prior work. The general form (Eq. 2) is presented as a unifying summary of known TDB models, but the regularization and the upper-bound analysis do not depend on renaming a known result as a new prediction. A separate mathematical concern exists: the step in Appendix B replacing (sum of trace norms)^(alpha/2) by the sum of the alpha/2-th powers is invalid for alpha/2 > 1, and Table 6 uses alpha > 2 for most models. That is a correctness or soundness issue in the proof, not a circularity: the claimed bound does not reduce by definition to its own inputs, and the counterexample does not show that any fitted parameter was renamed as a prediction. Therefore, under the circularity definition used here, the paper is self-contained against external benchmarks and receives a score of 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central derivation rests on an external matrix trace-norm lemma, plus a false power-inequality assumption that invalidates the upper-bound theorem for alpha>2. The tightness construction in Proposition 1 also mismatches tensor dimensions unless the n-rank equals P. The practical method depends on several tuned hyperparameters (alpha, lambda1, lambda2, D, P) rather than on invented physical entities.

free parameters (4)

alpha (power coefficient) = 3.0 on WN18RR for CP, ComplEx, SimplE, ANALOGY, QuatE; 2.25 on FB15k-237; 2.0 for TuckER; see Table 6
Tuned on validation MRR with Hyperopt; the theoretical bound only holds for alpha=2, yet most experiments use alpha>2.
lambda1 = lambda3 = e.g., 0.05 for CP on WN18RR, 0.005 for ComplEx on FB15k-237, 0.01 for TuckER on WN18RR; see Table 6
Searched over {0.001, 0.003, 0.005, 0.007, 0.01, 0.03, 0.05, 0.07}; the paper sets lambda1=lambda3 for all models even though the equality justification in Appendix C is only derived for CP and ComplEx.
lambda2 = lambda4 = e.g., 0.07 for CP on WN18RR, 0.05 for ComplEx on FB15k-237, 0.03 for TuckER on WN18RR; see Table 6
Tuned on validation MRR; the paper sets lambda2=lambda4 for all models to reduce the hyperparameter count.
embedding dimension D and number of parts P = D=2000/4000/256, P=1/2/4/256 depending on model and dataset; see Table 5
Model capacity choices that determine the general form Eq.(2); P controls expressiveness and computational cost.

assumptions (4)

standard math Lemma 1 from Ciliberto et al. (2017): ||Z||_*^alpha = min_{Z=UV^T} 1/2 (lambda||U||_F^{2alpha} + (1/lambda)||V||_F^{2alpha}).
Used as the foundation for Propositions 1 and 2; it expresses the trace norm as a minimum over factor norms.
ad hoc to paper The paper implicitly assumes ||sum_d A_d||_*^{alpha/2} <= sum_d ||A_d||_*^{alpha/2} for alpha/2 >= 1.
This is not a valid inequality for alpha>2 and breaks the proof of Proposition 1; a 1x1x1 CP counterexample with alpha=4 gives 24 <= 12.
ad hoc to paper In Proposition 1's tightness construction, H:1: is set to an n1 x r1 SVD factor while the general form requires H:d: to be n1 x P, and W is set to r1 x r2 x r3 while it should be P x P x P.
The construction is only consistent if r1=r2=r3=D, which is not stated; the proof silently assumes the n-rank equals the embedding dimension.
domain assumption Low overlapped trace norm L(X;alpha) corresponds to high entity and relation correlation and reduces overfitting.
The paper motivates minimizing L(X;alpha) as a surrogate for low-rank structure and correlation; this is a heuristic assumption, not derived.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Knowledge Graph Completion by Intermediate Variables Regularization." pith.science (2026). https://pith.science/paper/ST2MFSZ5

@misc{pith2026250602749,
  author       = {Pith},
  title        = {Pith review of: Knowledge Graph Completion by Intermediate Variables Regularization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ST2MFSZ5}},
  note         = {Machine review of arXiv:2506.02749}
}

read the original abstract

Knowledge graph completion (KGC) can be framed as a 3-order binary tensor completion task. Tensor decomposition-based (TDB) models have demonstrated strong performance in KGC. In this paper, we provide a summary of existing TDB models and derive a general form for them, serving as a foundation for further exploration of TDB models. Despite the expressiveness of TDB models, they are prone to overfitting. Existing regularization methods merely minimize the norms of embeddings to regularize the model, leading to suboptimal performance. Therefore, we propose a novel regularization method for TDB models that addresses this limitation. The regularization is applicable to most TDB models and ensures tractable computation. Our method minimizes the norms of intermediate variables involved in the different ways of computing the predicted tensor. To support our regularization method, we provide a theoretical analysis that proves its effect in promoting low trace norm of the predicted tensor to reduce overfitting. Finally, we conduct experiments to verify the effectiveness of our regularization technique as well as the reliability of our theoretical analysis. The code is available at https://github.com/changyi7231/IVR.

Figures

Figures reproduced from arXiv: 2506.02749 by the authors.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 25 canonical work pages

[1]

Templates for the solution of algebraic eigenvalue problems: a practical guide

Zhaojun Bai, James Demmel, Jack Dongarra, Axel Ruhe, and Henk van der Vorst. Templates for the solution of algebraic eigenvalue problems: a practical guide. SIAM, 2000

work page 2000
[2]

Tucker: Tensor factorization for knowledge graph completion

Ivana Bala z evi \'c , Carl Allen, and Timothy Hospedales. Tucker: Tensor factorization for knowledge graph completion. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5185--5194, 2019

work page 2019
[3]

Algorithms for hyper-parameter optimization

James Bergstra, R \'e mi Bardenet, Yoshua Bengio, and Bal \'a zs K \'e gl. Algorithms for hyper-parameter optimization. Advances in neural information processing systems, 24, 2011

2011
[4]

Translating embeddings for modeling multi-relational data

Antoine Bordes, Nicolas Usunier, Alberto Garcia-Duran, Jason Weston, and Oksana Yakhnenko. Translating embeddings for modeling multi-relational data. Advances in neural information processing systems, 26, 2013

work page 2013
[5]

Reexamining Low Rank Matrix Factorization for Trace Norm Regularization

Carlo Ciliberto, Dimitris Stamos, and Massimiliano Pontil. Reexamining low rank matrix factorization for trace norm regularization. arXiv preprint arXiv:1706.08934, 2017

work page Pith review arXiv 2017
[6]

Decompositions of a higher-order tensor in block terms—part i: Lemmas for partitioned matrices

Lieven De Lathauwer. Decompositions of a higher-order tensor in block terms—part i: Lemmas for partitioned matrices. SIAM Journal on Matrix Analysis and Applications, 30 0 (3): 0 1022--1032, 2008

work page 2008
[7]

Convolutional 2d knowledge graph embeddings

Tim Dettmers, Pasquale Minervini, Pontus Stenetorp, and Sebastian Riedel. Convolutional 2d knowledge graph embeddings. In Thirty-second AAAI conference on artificial intelligence, 2018

work page 2018
[8]

Adaptive subgradient methods for online learning and stochastic optimization

John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning and stochastic optimization. Journal of machine learning research, 12 0 (7), 2011

2011

Show all 29 references

[9]

Robust low-rank tensor recovery: Models and algorithms

Donald Goldfarb and Zhiwei Qin. Robust low-rank tensor recovery: Models and algorithms. SIAM Journal on Matrix Analysis and Applications, 35 0 (1): 0 225--253, 2014

2014
[10]

The expression of a tensor or a polyadic as a sum of products

Frank L Hitchcock. The expression of a tensor or a polyadic as a sum of products. Journal of Mathematics and Physics, 6 0 (1-4): 0 164--189, 1927

1927
[11]

A survey on knowledge graphs: Representation, acquisition, and applications

Shaoxiong Ji, Shirui Pan, Erik Cambria, Pekka Marttinen, and S Yu Philip. A survey on knowledge graphs: Representation, acquisition, and applications. IEEE transactions on neural networks and learning systems, 33 0 (2): 0 494--514, 2021

2021
[12]

Simple embedding for link prediction in knowledge graphs

Seyed Mehran Kazemi and David Poole. Simple embedding for link prediction in knowledge graphs. In Proceedings of the 32nd International Conference on Neural Information Processing Systems, pages 4289--4300, 2018

2018
[13]

Statistical predicate invention

Stanley Kok and Pedro Domingos. Statistical predicate invention. In Proceedings of the 24th international conference on Machine learning, pages 433--440, 2007

2007
[14]

Tensor decompositions and applications

Tamara G Kolda and Brett W Bader. Tensor decompositions and applications. SIAM review, 51 0 (3): 0 455--500, 2009

2009
[15]

Canonical tensor decomposition for knowledge base completion

Timoth \'e e Lacroix, Nicolas Usunier, and Guillaume Obozinski. Canonical tensor decomposition for knowledge base completion. In International Conference on Machine Learning, pages 2863--2872. PMLR, 2018

2018
[16]

Analogical inference for multi-relational embeddings

Hanxiao Liu, Yuexin Wu, and Yiming Yang. Analogical inference for multi-relational embeddings. In International conference on machine learning, pages 2168--2178. PMLR, 2017

2017
[17]

Tensor robust principal component analysis: Exact recovery of corrupted low-rank tensors via convex optimization

Canyi Lu, Jiashi Feng, Yudong Chen, Wei Liu, Zhouchen Lin, and Shuicheng Yan. Tensor robust principal component analysis: Exact recovery of corrupted low-rank tensors via convex optimization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pag...

2016
[18]

Square deal: Lower bounds and improved relaxations for tensor recovery

Cun Mu, Bo Huang, John Wright, and Donald Goldfarb. Square deal: Lower bounds and improved relaxations for tensor recovery. In International conference on machine learning, pages 73--81. PMLR, 2014

2014
[19]

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

Maximilian Nickel, Volker Tresp, and Hans-Peter Kriegel. A three-way model for collective learning on multi-relational data. In Proceedings of the 28th International Conference on International Conference on Machine Learning, pages 809--816, 2011

2011
[20]

Holographic embeddings of knowledge graphs

Maximilian Nickel, Lorenzo Rosasco, and Tomaso Poggio. Holographic embeddings of knowledge graphs. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 30, 2016

2016
[21]

Tensor completion algorithms in big data analytics

Qingquan Song, Hancheng Ge, James Caverlee, and Xia Hu. Tensor completion algorithms in big data analytics. ACM Transactions on Knowledge Discovery from Data (TKDD), 13 0 (1): 0 1--48, 2019

2019
[22]

Statistical performance of convex tensor decomposition

Ryota Tomioka, Taiji Suzuki, Kohei Hayashi, and Hisashi Kashima. Statistical performance of convex tensor decomposition. Advances in neural information processing systems, 24, 2011

2011
[23]

Representing text for joint embedding of text and knowledge bases

Kristina Toutanova, Danqi Chen, Patrick Pantel, Hoifung Poon, Pallavi Choudhury, and Michael Gamon. Representing text for joint embedding of text and knowledge bases. In Proceedings of the 2015 conference on empirical methods in natural language processing, pages 1499--1509, 2015

2015
[24]

Knowledge graph completion via complex tensor factorization

Th \'e o Trouillon, Christopher R Dance, \'E ric Gaussier, Johannes Welbl, Sebastian Riedel, and Guillaume Bouchard. Knowledge graph completion via complex tensor factorization. Journal of Machine Learning Research, 18: 0 1--38, 2017

2017
[25]

Some mathematical notes on three-mode factor analysis

Ledyard R Tucker. Some mathematical notes on three-mode factor analysis. Psychometrika, 31 0 (3): 0 279--311, 1966

1966
[26]

Embedding entities and relations for learning and inference in knowledge bases

Bishan Yang, Wen-tau Yih, Xiaodong He, Jianfeng Gao, and Li Deng. Embedding entities and relations for learning and inference in knowledge bases. arXiv e-prints, pages arXiv--1412, 2014

2014
[27]

Neural, symbolic and neural-symbolic reasoning on knowledge graphs

Jing Zhang, Bo Chen, Lingxi Zhang, Xirui Ke, and Haipeng Ding. Neural, symbolic and neural-symbolic reasoning on knowledge graphs. AI Open, 2: 0 14--35, 2021

2021
[28]

Quaternion knowledge graph embeddings

Shuai Zhang, Yi Tay, Lina Yao, and Qi Liu. Quaternion knowledge graph embeddings. In Proceedings of the 33rd International Conference on Neural Information Processing Systems, pages 2735--2745, 2019

2019
[29]

Duality-induced regularizer for tensor factorization based knowledge graph completion

Zhanqiu Zhang, Jianyu Cai, and Jie Wang. Duality-induced regularizer for tensor factorization based knowledge graph completion. Advances in Neural Information Processing Systems, 33, 2020

2020

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Templates for the solution of algebraic eigenvalue problems: a practical guide

Zhaojun Bai, James Demmel, Jack Dongarra, Axel Ruhe, and Henk van der Vorst. Templates for the solution of algebraic eigenvalue problems: a practical guide. SIAM, 2000

work page 2000

[2] [2]

Tucker: Tensor factorization for knowledge graph completion

Ivana Bala z evi \'c , Carl Allen, and Timothy Hospedales. Tucker: Tensor factorization for knowledge graph completion. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5185--5194, 2019

work page 2019

[3] [3]

Algorithms for hyper-parameter optimization

James Bergstra, R \'e mi Bardenet, Yoshua Bengio, and Bal \'a zs K \'e gl. Algorithms for hyper-parameter optimization. Advances in neural information processing systems, 24, 2011

2011

[4] [4]

Translating embeddings for modeling multi-relational data

Antoine Bordes, Nicolas Usunier, Alberto Garcia-Duran, Jason Weston, and Oksana Yakhnenko. Translating embeddings for modeling multi-relational data. Advances in neural information processing systems, 26, 2013

work page 2013

[5] [5]

Reexamining Low Rank Matrix Factorization for Trace Norm Regularization

Carlo Ciliberto, Dimitris Stamos, and Massimiliano Pontil. Reexamining low rank matrix factorization for trace norm regularization. arXiv preprint arXiv:1706.08934, 2017

work page Pith review arXiv 2017

[6] [6]

Decompositions of a higher-order tensor in block terms—part i: Lemmas for partitioned matrices

Lieven De Lathauwer. Decompositions of a higher-order tensor in block terms—part i: Lemmas for partitioned matrices. SIAM Journal on Matrix Analysis and Applications, 30 0 (3): 0 1022--1032, 2008

work page 2008

[7] [7]

Convolutional 2d knowledge graph embeddings

Tim Dettmers, Pasquale Minervini, Pontus Stenetorp, and Sebastian Riedel. Convolutional 2d knowledge graph embeddings. In Thirty-second AAAI conference on artificial intelligence, 2018

work page 2018

[8] [8]

Adaptive subgradient methods for online learning and stochastic optimization

John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning and stochastic optimization. Journal of machine learning research, 12 0 (7), 2011

2011