Block Randomized Optimization for Adaptive Hypergraph Learning

Constantine Kotropoulos; Georgios Karantaidis; Ioannis Sarridis

REVIEW 3 major objections 6 minor 25 references

Block Randomized Optimization for Adaptive Hypergraph Learning

T0 review · 3 major / 6 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read The paper claims that replacing exact matrix inversion with conjugate gradient cuts the computational time of adaptive hypergraph weight estimation by about 92% while keeping or slightly improving F1-based image tagging.

desk verdict The conjugate-gradient half is a plausible incremental speedup; the block randomized SVD half is mathematically invalid as written, which guts the paper's central claim. read the letter →

arxiv 1908.08281 v1 pith:MD7J33V5 submitted 2019-08-22 cs.SI stat.ML

Georgios Karantaidis , Ioannis Sarridis , Constantine Kotropoulos This is my paper · ORCID

classification cs.SIstat.ML

keywords hypergraphlearningrandomizedSVDconjugategradientimagetaggingadaptiveweightestimationlow-rankapproximationsocialmediamatrixinversion

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

This paper tries to show that the slow step in adaptive hypergraph learning—solving a large linear system to rank images by tags—can be accelerated by two replacements for exact matrix inversion: block randomized SVD and the conjugate gradient method. The authors test both inside an existing adaptive weight-estimation scheme for image tagging on a Flickr-derived dataset. They report that conjugate gradient cuts total computation from 8,530 seconds to 727 seconds, about 92%, while F1@1 edges up from 0.425 to 0.427, and that block randomized SVD yields a 46% reduction with nearly unchanged F1. If these numbers hold, the payoff is that hypergraph models of high-order relations can be applied to much larger image collections and to near-real-time single-query tagging.

What carries the argument

The central object is the matrix $X = I - \frac{1}{1+\theta}A$, where $A = D_v^{-1/2} H W D_e^{-1} H^\top D_v^{-1/2}$ is the normalized hypergraph similarity matrix built from the incidence matrix $H$ and hyperedge weights $W$; the ranking vector is $f^* = \frac{\theta}{1+\theta} X^{-1} y$. The first acceleration uses a $2\times 2$ block tessellation of $X$ with $X_{11}$ and $X_{22}$ treated as low-rank diagonal blocks, inverted via randomized SVD with subspace iteration and combined through the Schur-complement block inversion formula. The second acceleration replaces inversion entirely by solving $X f = \frac{\theta}{1+\theta} y$ with the conjugate gradient iterations given in the paper.

What would settle it

Compute the numerical rank and smallest singular values of the diagonal blocks $X_{11}$ and $X_{22}$ on the paper's hypergraph; if any block has smallest singular value near machine precision relative to its largest, the block randomized inversion used to compute $X^{-1}$ is not justified.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central claim is that adaptive hyperedge weight estimation in hypergraph learning does not require exact inversion of the matrix $X = I - \frac{1}{1+\theta} A$. The paper proposes two fast solvers for $X f = \frac{\theta}{1+\theta} y$: tessellating $X$ into diagonal blocks and inverting each block by randomized SVD using the Schur-complement block inversion formula, or solving the system directly with conjugate gradient iterations. On the dataset used, both preserve the F1 measure at ranks 1, 2, 5, and 10; conjugate gradient reduces runtime from 8,530 to 727 seconds for the adaptive scheme and from 4,267 to 355 seconds for the fixed-weight scheme, while F1@1 is 0.427 versus 0.425 in the adaptive case.

Load-bearing premise

The block randomized SVD half assumes the diagonal blocks $X_{11}$ and $X_{22}$ are low-rank yet invertible enough for the block inversion formula to be accurate, but a low-rank matrix is singular and the paper gives no proof or numerical evidence of invertibility.

Editorial extensions

If this is right

The adaptive update of hyperedge weights becomes cheap enough to keep in the loop, so users get the improved F1 of adaptive weights instead of settling for fixed weights.
A single query image can be tagged in about 0.56 seconds with the conjugate-gradient adaptive method, which is fast enough for interactive or real-time applications.
Both the fixed-weight and adaptive-weight ranking formulations solve the same linear system, so the conjugate-gradient speedup applies to both settings.
The near-linear growth of runtime with the number of images is no longer prohibitive, making larger hypergraphs than the 1,292-image test set feasible.

Reading between the lines

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

The conjugate-gradient result stands independently of the block randomized SVD construction; even if the low-rank block inversion were not valid, the reported speedup would still be available because CG only needs matrix-vector products with $X$.
The block randomized SVD approach could be made rigorous by adding a pseudoinverse or perturbation analysis for the Schur complements; without that, its 46% time reduction is an empirical observation rather than a fully derived method.
The same hypergraph Laplacian linear system appears in other high-order relation tasks such as recommendation and geo-location prediction, so the speedup logic likely transfers, though the paper only demonstrates it for image tagging.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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

Colleague,

Here's my read on Karantaidis et al. The paper offers two speedups for adaptive hypergraph weight estimation: block randomized SVD to invert X, and conjugate gradient to solve X f = y. The CG part is straightforward but sound: they replace a direct solve with CG, report a 92% time reduction on their dataset with unchanged F1, and the numbers are plausible. It's a minor but legitimate incremental result.

The block randomized SVD part is another story. The derivation in Section 3 uses the block inversion formula (12) on a partition where X11 and X22 are explicitly described as low-rank and rank-deficient. A square low-rank matrix is singular and has no inverse, so the formula as stated is invalid. They then 'invert' those blocks via randomized SVD, but a truncated SVD yields a pseudoinverse, not an inverse, and discarding small singular values is particularly dangerous for inversion because the inverse is dominated by them. There is no pseudoinverse correction, no regularization, and no error bound. Equation (15), which forms X^{-1} from a rank-50 truncation, compounds the problem. The BR-ITH and BR-ITH-HWEG rows in Tables 2 and 3 are therefore not produced by a valid algorithm as described. This is an internal mathematical inconsistency, not a modeling disagreement.

Other soft spots: the F1 comparison is essentially a consistency check—an iterative solver for the same linear system should yield nearly the same rankings, so near-equal F1 carries little predictive weight. The dataset is small (1,292 images), there are no error bars on the single-run timing or F1 numbers, no code or data are released, and the rank thresholds (50, then 500) look arbitrary. The abstract's claim that both approaches work is only true for CG.

Who is this for? Someone looking to scale hypergraph learning might find the CG speedup worth reimplementing, but they'd need to verify it on their own data. As a paper, the central claim fails because half of it is invalid. A serious referee would almost certainly reject it in its current form. If the BR section were removed or fixed, the CG-only version could be a modest workshop-level contribution.

My recommendation: desk reject, or send back only if the author is willing to drop the BR approach entirely and reframe the paper as a CG speedup with proper error analysis. As is, I wouldn't spend referee time on it.

Best, [Your name]

Referee Report

3 major / 6 minor

Summary. The paper proposes two computational accelerations for adaptive hypergraph weight estimation in image tagging. In the first approach (BR-ITH / BR-ITH-HWEG), the full-rank matrix X is tessellated into blocks, the diagonal blocks are declared low-rank, and block randomized SVD is used to invert X in the ranking-vector update. In the second approach (CG-ITH / CG-ITH-HWEG), the same linear system is solved by conjugate gradient. Experiments on a Flickr-derived dataset report large time reductions for both variants (e.g., 8,530 to 727 seconds for CG-ITH-HWEG) with F1 scores at the same level as the baseline.

Significance. If the claims were fully established, the conjugate-gradient variant would be a practically useful speedup for a standard hypergraph learning pipeline, and the time-reduction numbers are striking. The paper deserves credit for a clear experimental comparison against its own baselines and for reporting per-image timings. However, the block randomized SVD variant rests on an invalid mathematical step, and the F1 comparisons are not a strong substitute for an accuracy check of the approximate inverse. The paper therefore does not establish its headline that both proposed approaches are accurate and fast.

major comments (3)

[Section 3, after Eq. (11)] The derivation of the block randomized inverse is internally inconsistent. The text states that X11 and X22 are low-rank submatrices, and the Introduction states that creating rank deficient blocks in the main diagonal is intended. A square low-rank (rank-deficient) matrix is singular, but Eq. (12) requires X11^{-1} and X22^{-1}. If 'low-rank' is meant as 'numerical low-rank,' the requested inverse is at best ill-conditioned, and no pseudoinverse, regularization, or error bound is provided. Because Eqs. (12)-(14) are the basis for the BR-ITH and BR-ITH-HWEG rows in Tables 2 and 3, those results are not produced by a valid algorithm as described.
[Section 3, Eq. (15)] The use of a truncated SVD to form X^{-1} is not justified. Eq. (15) writes X^{-1} = V Σ^{-1} U^T, but if Σ retains only the leading rank-50 singular values, the matrix V Σ^{-1} U^T is a pseudoinverse-like low-rank object, not the inverse of X. Discarding small singular values is particularly dangerous for inversion, because the inverse is dominated by the smallest retained singular values in exact arithmetic. Without a bound on ||X^{-1} - V Σ^{-1} U^T|| or on the resulting error in f*, no claim about solving Eq. (3) can be made.
[Section 4, Tables 2 and 3] The experimental validation is too weak to support the 'same performance' claim. Since CG and BR are intended to solve the same linear system, near-identical F1 scores are the expected outcome of a consistency check rather than independent evidence of correctness. The paper should report convergence residuals, numbers of CG iterations, actual low-rank block ranks and approximation errors, and ideally timing breakdowns; without these, the F1 comparison cannot distinguish an accurate fast solver from an inaccurate one that happens to rank the test images similarly.

minor comments (6)

[Section 2, Eq. (1)] The hyperedge degree matrix is denoted Dv twice; the second occurrence should be De.
[Introduction] The text contains spelling errors, including 'matrix tesselation' and 'first appoach.'
[Section 4] The method name alternates between 'ITH-HWEG' and 'ITH-WHEG'; the latter appears to be a typo.
[Section 4, Figure 1] Figure 1 is referenced but its axes and legend are not described in the text, so the reader cannot tell what is plotted for each method.
[Section 3, nested tessellation] The choice of the minimum diagonal-block rank (50, then 500) is ad hoc, and no sensitivity analysis is reported.
[General] No code or dataset release is mentioned, which limits reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the F1 comparisons are internal consistency checks of an approximate solver, and all load-bearing formulas come from external sources.

full rationale

The paper's stated goal is to reduce the computational cost of an existing adaptive hypergraph-learning loop by replacing the exact solve in Eq. (3) with a block randomized SVD inversion or a conjugate-gradient solve of Eq. (6), and then to verify that the F1 measures are not degraded. Since the conjugate-gradient and randomized methods are intended to approximate the same linear system that ITH and ITH-HWEG solve, near-identical F1 values are the expected outcome of a consistency check, not an independent prediction; this is not a circular derivation. The matrix-inversion and randomized-SVD ingredients are cited to external sources (Hackbusch [14], Halko et al. [2]), and the conjugate-gradient recurrences are cited to [20]; no uniqueness theorem or central premise is imported from the authors' own prior work. The dataset and ITH/ITH-HWEG baselines come from the authors' previous papers [12,13,15], but those are used as experimental comparison points rather than as evidence for the new contribution, and the time measurements in Table 2 are concrete empirical results. The paper's weakness is a mathematical inconsistency in calling the diagonal blocks X11 and X22 low-rank and then inverting them in Eq. (12), since a rank-deficient square matrix is singular; however, that is an internal validity or error issue, not a circularity of the kind where a claimed result equals its inputs by construction. Accordingly, the circularity score is 0.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests on standard randomized SVD and CG machinery, plus one ad hoc structural assumption: that tessellation produces low-rank, invertible diagonal blocks. The free parameters are the block rank thresholds, the oversampling parameter, and the CG stopping rule, none of which are justified or reported, making the headline numbers hard to reproduce.

free parameters (4)

Minimum diagonal block rank, first update = 50
Hand-chosen threshold for nested tessellation before weight updates; no sensitivity analysis is given.
Minimum diagonal block rank, second update = 500
Hand-chosen threshold after the weight update, justified by X being sparser; no justification for the value.
Randomized SVD oversampling parameter pi = not reported
The paper notes 5 or 10 is typical [2], but does not state what was used, and the approximation error depends on it.
Conjugate gradient stopping criterion = not reported
Number of iterations or residual tolerance is not given, so the reported 92% time reduction is tied to unspecified convergence settings.

assumptions (5)

standard math Randomized SVD provides a reliable low-rank approximation for matrices with low numerical rank, per Halko et al. [2].
The method inherits these guarantees, but the paper does not verify that the diagonal blocks meet the low-rank condition.
standard math The block inversion formula (12) is valid when X11, X22, and the Schur complements Z1, Z2 are invertible.
Used in Eq. (12); the paper does not check invertibility of the low-rank blocks.
domain assumption Zhou's hypergraph ranking formula f* = theta/(1+theta) X^{-1} y correctly ranks tags in this application.
Taken from [18,19]; central to the optimization problem being solved.
ad hoc to paper Tessellating X yields diagonal blocks that are low-rank, so randomized SVD can invert them accurately.
This is the key unproven premise of the block randomized SVD approach; the paper provides no evidence for it.
domain assumption The dataset and preprocessing from [12,13] are appropriate and yield the same experimental setup.
The paper reuses the prior dataset without reproducing preprocessing details.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Block Randomized Optimization for Adaptive Hypergraph Learning." pith.science (2026). https://pith.science/paper/MD7J33V5

@misc{pith2026190808281,
  author       = {Pith},
  title        = {Pith review of: Block Randomized Optimization for Adaptive Hypergraph Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MD7J33V5}},
  note         = {Machine review of arXiv:1908.08281}
}

read the original abstract

The high-order relations between the content in social media sharing platforms are frequently modeled by a hypergraph. Either hypergraph Laplacian matrix or the adjacency matrix is a big matrix. Randomized algorithms are used for low-rank factorizations in order to approximately decompose and eventually invert such big matrices fast. Here, block randomized Singular Value Decomposition (SVD) via subspace iteration is integrated within adaptive hypergraph weight estimation for image tagging, as a first approach. Specifically, creating low-rank submatrices along the main diagonal by tessellation permits fast matrix inversions via randomized SVD. Moreover, a second approach is proposed for solving the linear system in the optimization problem of hypergraph learning by employing the conjugate gradient method. Both proposed approaches achieve high accuracy in image tagging measured by F1 score and succeed to reduce the computational requirements of adaptive hypergraph weight estimation.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 25 canonical work pages

[1]

Block Randomized Optimization for Adaptive Hypergraph Learning

INTRODUCTION Social media platforms store huge amount of multimedia content daily and encourage users to provide descriptions and tags about it. This perpetual, dynamic procedure peaks with users sharing the con- tent among the community. As a consequence, ever increasing data are stored every second in companies’ servers. Handling these data becomes a ve...

work page Pith review arXiv 1908
[2]

ADAPTIVE HYPEREDGE WEIGHT UPDA TING MODEL Let|·| denote set cardinality,||·|| be thel2-norm of a vector, and I denote the identity matrix of compatible dimensions. A hyper- graph G(V,E,w ) captures high-order relationships in social me- dia, where V is the set of vertices, E is the set of hyperedges and w() is real-valued function assigning weights in hyp...

work page
[3]

The ﬁrst stage comprises random sampling in order to ﬁnd a lower- dimensional subspace which captures the most of the action of X∈ Rm×n

BLOCK RANDOMIZED SVD AND CONJUGA TE GRADIENT METHOD FOR OPTIMIZA TION Randomized matrix approximations are performed in two stages. The ﬁrst stage comprises random sampling in order to ﬁnd a lower- dimensional subspace which captures the most of the action of X∈ Rm×n. To do so a column orthonormal matrixS∈ Rm×l should be computed such that ||X− SS⊤X||F ⩽ϵ...

work page
[4]

It contains a large amount of Greek places of interest along with valuable information related to them

DA TASET DESCRIPTION AND EXPERIMENTS The same dataset used in [12, 13] is employed here, retaining the same experimental setup. It contains a large amount of Greek places of interest along with valuable information related to them. In par- ticular, geotagged photos, both indoor and outdoor, are accompanied with auxiliary information, such as id, title, ow...

work page
[5]

These methods are block randomized SVD for matrix inversion and conjugate gradient for solving a set of linear equations both related to optimizing f given ﬁxed hyperedge weights w

CONCLUSIONS AND FUTURE WORK Two different approaches for optimizing the ranking vector in hy- pergraph learning have been proposed. These methods are block randomized SVD for matrix inversion and conjugate gradient for solving a set of linear equations both related to optimizing f given ﬁxed hyperedge weights w. It was shown that both approaches yield com...

work page
[6]

Multimedia social search based on hy- pergraph learning,

C. Kotropoulos, “Multimedia social search based on hy- pergraph learning,” in Graph-Based Social Media Analysis , I. Pitas, Ed., vol. 39, pp. 215–273. CRC Press, 2016

work page 2016
[7]

Finding structure with randomness: Probabilistic algorithms for constructing ap- proximate matrix decompositions,

N. Halko, P. Martinsson, and J. A Tropp, “Finding structure with randomness: Probabilistic algorithms for constructing ap- proximate matrix decompositions,” SIAM Review, vol. 53, no. 2, pp. 217–288, 2011

work page 2011
[8]

Low-rank approximation of large-scale matrices via randomized methods,

S. Hatamirad and M. M. Pedram, “Low-rank approximation of large-scale matrices via randomized methods,” J. of Super- computing, vol. 74, no. 2, pp. 830–844, 2018

work page 2018

Show all 25 references

[9]

A randomized al- gorithm for the decomposition of matrices,

P. Martinsson, V . Rokhlin, and M. Tygert, “A randomized al- gorithm for the decomposition of matrices,”Applied and Com- putat. Harmonic Analysis, vol. 30, no. 1, pp. 47–68, 2011

2011
[10]

Randomized algorithms for the low-rank approx- imation of matrices,

E. Liberty, F. Woolfe, P. Martinsson, V . Rokhlin, and M. Tygert, “Randomized algorithms for the low-rank approx- imation of matrices,” Proc. of the National Academy of Sci- ences, vol. 104, no. 51, pp. 20167–20172, 2007

2007
[11]

A fast frequent directions algorithm for low rank approximation,

D. Teng and D. Chu, “A fast frequent directions algorithm for low rank approximation,” IEEE Trans. Pattern Analysis and Machine Intelligence, 2018, to appear

2018
[12]

Low rank approxi- mation of a sparse matrix based on lu factorization with column and row tournament pivoting,

L. Grigori, S. Cayrols, and J. W Demmel, “Low rank approxi- mation of a sparse matrix based on lu factorization with column and row tournament pivoting,” SIAM J. Scientiﬁc Computing, vol. 40, no. 2, pp. 181–209, 2018

2018
[13]

Fast randomized singular value thresholding for low-rank optimization,

T. Oh, Y . Matsushita, Y . Tai, and I. Kweon, “Fast randomized singular value thresholding for low-rank optimization,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 40, no. 2, pp. 376–391, 2018

2018
[14]

Image ranking via attribute boosted hypergraph,

Z. Yu, S. Tang, Y . Zhang, and J. Shao, “Image ranking via attribute boosted hypergraph,” in Proc. 13th Paciﬁc-Rim Conf. Advances Multimedia Inf. Process., 2012, pp. 779–789

2012
[15]

Uniﬁed hy- pergraph for image ranking in a multimodal context,

J. Xu, V . Singh, Z. Guan, and B. S. Manjunath, “Uniﬁed hy- pergraph for image ranking in a multimodal context,” in Proc. IEEE Int. Conf. Acoustics, Speech, and Signal Process., 2012, pp. 2333–2336

2012
[16]

Visual-textual joint relevance learning for tag-based social image search,

Y . Gao, M. Wang, Z. J. Zha, J. Shen, X. Li, and X. Wu, “Visual-textual joint relevance learning for tag-based social image search,” IEEE Trans. Image Process. , vol. 22, no. 1, pp. 363–376, 2013

2013
[17]

Weight estimation in hyper- graph learning,

K. Pliakos and C. Kotropoulos, “Weight estimation in hyper- graph learning,” in Proc. IEEE Int. Conf. Acoustics, Speech, and Signal Process., 2015, pp. 1161–1165

2015
[18]

Adaptive al- gorithms for hypergraph learning,

A. Chasapi, C. Kotropoulos, and K. Pliakos, “Adaptive al- gorithms for hypergraph learning,” in Proc. IEEE Int. Conf. Acoustics, Speech and Signal Process., 2016, pp. 1179–1183

2016
[19]

Hackbusch, A Sparse Matrix Arithmetic Based on H- Matrices

W. Hackbusch, A Sparse Matrix Arithmetic Based on H- Matrices. Part I: Introduction to H-Matrices, vol. 62, Springer, 1999

1999
[20]

Simultaneous image tagging and geo-location prediction within hypergraph ranking frame- work,

K. Pliakos and C. Kotropoulos, “Simultaneous image tagging and geo-location prediction within hypergraph ranking frame- work,” in Proc. IEEE Int. Conf. Acoustics, Speech, and Signal Process., 2014, pp. 6944–6948

2014
[21]

Learning with hy- pergraphs: Clustering, classiﬁcation, and embedding,

D. Zhou, J. Huang, and B. Sch ¨olkopf, “Learning with hy- pergraphs: Clustering, classiﬁcation, and embedding,” in Ad- vances Neural Inf. Process. Systems, 2007, vol. 19, pp. 1601– 1608

2007
[22]

Higher order learn- ing with graphs,

S. Agarwal, K. Branson, and S. Belongie, “Higher order learn- ing with graphs,” in Proc. 23rd Int. Conf. Machine Learning, 2006, pp. 17–24

2006
[23]

Learning with local and global consistency,

D. Zhou, O. Bousquet, T. N. Lal, J. Weston, and B. Sch ¨olkopf, “Learning with local and global consistency,” in Advances Neural Inf. Process. Systems, 2004, vol. 16, pp. 321–328

2004
[24]

Music recommendation by uniﬁed hypergraph: Combining social media information and music content,

J. Bu, S. Tan, C. Chen, C. Wang, H. Wu, Z. Lijun, and X. He, “Music recommendation by uniﬁed hypergraph: Combining social media information and music content,” in Proc. ACM Conf. Multim., 2010, pp. 391–400

2010
[25]

Barrett, M

R. Barrett, M. Berry, T. F. Chan, J. Demmel, J. Donato, J. Don- garra, V . Eijkhout, R. Pozo, C. Romine, and H. Van der V orst, Templates for the Solution of Linear Systems: Building Blocks for Iterative Methods, SIAM, Philadelphia, 1994

1994

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [1]

Block Randomized Optimization for Adaptive Hypergraph Learning

INTRODUCTION Social media platforms store huge amount of multimedia content daily and encourage users to provide descriptions and tags about it. This perpetual, dynamic procedure peaks with users sharing the con- tent among the community. As a consequence, ever increasing data are stored every second in companies’ servers. Handling these data becomes a ve...

work page Pith review arXiv 1908

[2] [2]

ADAPTIVE HYPEREDGE WEIGHT UPDA TING MODEL Let|·| denote set cardinality,||·|| be thel2-norm of a vector, and I denote the identity matrix of compatible dimensions. A hyper- graph G(V,E,w ) captures high-order relationships in social me- dia, where V is the set of vertices, E is the set of hyperedges and w() is real-valued function assigning weights in hyp...

work page

[3] [3]

The ﬁrst stage comprises random sampling in order to ﬁnd a lower- dimensional subspace which captures the most of the action of X∈ Rm×n

BLOCK RANDOMIZED SVD AND CONJUGA TE GRADIENT METHOD FOR OPTIMIZA TION Randomized matrix approximations are performed in two stages. The ﬁrst stage comprises random sampling in order to ﬁnd a lower- dimensional subspace which captures the most of the action of X∈ Rm×n. To do so a column orthonormal matrixS∈ Rm×l should be computed such that ||X− SS⊤X||F ⩽ϵ...

work page

[4] [4]

It contains a large amount of Greek places of interest along with valuable information related to them

DA TASET DESCRIPTION AND EXPERIMENTS The same dataset used in [12, 13] is employed here, retaining the same experimental setup. It contains a large amount of Greek places of interest along with valuable information related to them. In par- ticular, geotagged photos, both indoor and outdoor, are accompanied with auxiliary information, such as id, title, ow...

work page

[5] [5]

These methods are block randomized SVD for matrix inversion and conjugate gradient for solving a set of linear equations both related to optimizing f given ﬁxed hyperedge weights w

CONCLUSIONS AND FUTURE WORK Two different approaches for optimizing the ranking vector in hy- pergraph learning have been proposed. These methods are block randomized SVD for matrix inversion and conjugate gradient for solving a set of linear equations both related to optimizing f given ﬁxed hyperedge weights w. It was shown that both approaches yield com...

work page

[6] [6]

Multimedia social search based on hy- pergraph learning,

C. Kotropoulos, “Multimedia social search based on hy- pergraph learning,” in Graph-Based Social Media Analysis , I. Pitas, Ed., vol. 39, pp. 215–273. CRC Press, 2016

work page 2016

[7] [7]

Finding structure with randomness: Probabilistic algorithms for constructing ap- proximate matrix decompositions,

N. Halko, P. Martinsson, and J. A Tropp, “Finding structure with randomness: Probabilistic algorithms for constructing ap- proximate matrix decompositions,” SIAM Review, vol. 53, no. 2, pp. 217–288, 2011

work page 2011

[8] [8]

Low-rank approximation of large-scale matrices via randomized methods,

S. Hatamirad and M. M. Pedram, “Low-rank approximation of large-scale matrices via randomized methods,” J. of Super- computing, vol. 74, no. 2, pp. 830–844, 2018

work page 2018