Pith. sign in

REVIEW 4 major objections 5 minor 52 references

SGL-HNCS learns signed graphs from partially observed signals by explicitly correcting for hidden nodes, recovering positive and negative edge structure with higher accuracy than methods that assume full observability.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A signed graph learning method that explicitly models hidden nodes via low-rank, column-sparse correction matrices, solved with block coordinate descent.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection A useful new task and promising experiments, but the central optimization problem is unbounded below and the R^- update contradicts the stated objective, so the convergence claims do not hold as written. the 4 major comments →

arxiv 2509.09120 v1 pith:PVZZ6TPT submitted 2025-09-11 eess.SP

Signed Graph Learning with Hidden Nodes

Rong Ye , Xue-Qin Jiang , Hui Feng , Jian Wang , Runhe Qiu This is my paper
classification eess.SP
keywords signed graph learninghidden nodesgraph signal processingcolumn-sparsity regularizationgraph Laplacian estimationsmooth graph signalsblock coordinate descentADMM
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 tackles graph learning when only some nodes' signals are observed: it aims to recover the signed graph Laplacian on the observed nodes without pretending hidden nodes don't exist. It does so by decomposing the signed graph into positive and negative unsigned graphs, rewriting the smoothness objective so the hidden-node contribution appears as low-rank correction matrices, and regularizing those matrices with a group-Lasso (column-sparsity) penalty. The method, SGL-HNCS, solves the resulting constrained optimization with a block coordinate descent scheme whose inner Laplacian update uses ADMM; the authors prove convergence to stationary points. Experiments on synthetic and real-world data report that SGL-HNCS recovers the signed structure more accurately than signed-graph learning without hidden-node handling and than unsigned-graph methods with hidden nodes.

Core claim

The central claim is that the observable-node signed graph Laplacian can be estimated from the observed sample covariance even when hidden nodes influence the measurements, provided the hidden influence is weak enough to be modeled as a column-sparse correction. Starting from the smoothness assumption — signals are similar across positive edges and dissimilar across negative edges — the authors derive a block decomposition of the total variation in which the incomplete covariance only enters through tr(C_hat_B L_s_B) plus correction terms P~s and R_s. The key move is to replace the unobservable L_s_B with a proper combinatorial Laplacian L~s_B and to absorb the hidden-node degree matrix into

What carries the argument

The load-bearing object is the column-sparse correction matrix P~s = C_hat_B D~s_B/2 + C_hat_BH (L_s_BH)^T, which bundles the unobserved cross-covariance and the hidden-node degree correction. Its column sparsity — enforced by the group-Lasso norm ||P~s||_{2,1} — is what lets the estimator account for hidden nodes without observing them. The algorithm alternates three blocks: an ADMM update for the Laplacian pair (L~+_B, L~-_B) with a complementarity constraint separating positive and negative edges, and convex updates for P~s and R_s. The convergence proof relies on coercivity, Lipschitz differentiability, and regularity of the objective, invoking standard nonconvex ADMM and BCD convergence

Load-bearing premise

The whole construction rests on the assumption that the hidden-node correction matrix P~s is column-sparse, which holds only when hidden nodes are few and each connects sparsely to observed nodes; if hidden nodes are numerous or densely connected, the group-Lasso penalty mis-specifies the hidden influence and the recovered Laplacian is biased.

What would settle it

Run SGL-HNCS on a synthetic signed graph with a fixed set of observed nodes but with the number of hidden nodes pushed past the column-sparse regime (e.g., H approaching B or dense observed-hidden edges) and measure F-score/RelErr: if the method's advantage over scSGL disappears exactly when column sparsity fails, the claim is confirmed; if it still recovers accurately, the column-sparsity assumption is not the operative mechanism.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If graph signals are observed on only a subset of nodes, the observable-node signed Laplacian is identifiable under the column-sparse correction model, and ignoring hidden nodes biases the estimated edge signs and weights.
  • The decomposition of a signed graph into positive and negative unsigned graphs extends to the hidden-node setting, so both edge signs can be recovered simultaneously rather than treating the signed graph as one unsigned graph.
  • The BCD-ADMM algorithm comes with stationary-point convergence guarantees, and the convergence experiments show the objective error drops sharply within about ten iterations.
  • On real-world social/recommendation data, recovering both positive and negative edges while accounting for hidden nodes yields substantially higher F-score and NMI than methods that assume full observability.
  • The per-iteration cost is dominated by the ADMM inner update and scales polynomially in the number of observed nodes; the hidden-node block adds negligible cost when hidden nodes are few.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper's own experiments show SGL-HNCS and its low-rank variant SGL-HNLR perform almost identically, which suggests column sparsity may be acting as a proxy for low-rank structure in the small-hidden-node regime; a nuclear-norm-only formulation might be simpler without sacrificing accuracy.
  • Because the method treats the number of hidden nodes as known input, a natural extension is to estimate H itself, for example by model selection or rank detection on the correction matrix, rather than assuming it is given.
  • The block-decomposition trick for absorbing hidden-node influence is tied to the smoothness prior; the same decomposition could likely be adapted to other graph-signal priors such as stationarity, yielding signed-graph learning methods for hidden nodes beyond the smooth case.
  • If column sparsity is the active mechanism, the method's advantage should shrink as the number of hidden nodes grows or as observed-hidden connectivity becomes dense; testing this directly would separate the low-rank correction effect from the group-Lasso effect.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper addresses signed graph learning when only a subset of nodes is observed. It decomposes the signed graph into positive and negative unsigned graphs, writes the total variation in terms of observable-node covariances plus hidden-node correction matrices, and proposes the constrained optimization problem (6) with column-sparsity regularization on the correction terms. The solution algorithm is a block coordinate descent (BCD) method in which the Laplacian block is updated by an ADMM scheme and the correction and residual blocks by CVX. The authors claim convergence of both algorithms and report synthetic and real-data experiments showing improved accuracy over baselines that ignore hidden nodes.

Significance. The problem is relevant to graph signal processing, and the idea of modeling hidden-node influence through column-sparse correction matrices is reasonable. If the formulation were well-posed, the paper would fill a clear gap: previous hidden-node graph learning methods are unsigned, while signed graphs appear in social and recommendation networks. However, as written, the central optimization problem (6) is unbounded below, and the implemented R-update does not match it. These are not presentation issues; they invalidate the claimed equivalence between the stated problem and the algorithm, and they break the convergence proofs. The paper does not provide code or machine-checked derivations, and the convergence analysis relies on unverified assumptions about nonconvex projections and exact block minimization. The empirical comparisons are suggestive but currently disconnected from the stated objective. The central claim is therefore not established in the submitted form.

major comments (4)
  1. [§3.2, Eq. (6)] Problem (6) is unbounded below. The objective contains −tr(R^−), while the only constraints involving R^− are tr(C_B L~_B^−)+2tr(P~^−)+tr(R^−)≥0 and tr(R^−)≥0. For any feasible (L~,P~,R^+), set R^−=tI_H with t≥max(0,−tr(C_B L~_B^−)−2tr(P~^−)). The constraints are satisfied for all such t, and the objective goes to −∞ as t→∞. Hence (6) has no finite minimizer. This also invalidates the compact-level-set argument used in the proof of Theorem 2 (Section 5.2). The claim that Algorithm 2 solves (6) is unsupported.
  2. [§4.3, Eq. (18)] The R^s update in Eq. (18) minimizes tr(R^s) for s∈{+,−}, but in (6) the R^− block enters as −tr(R^−). The subproblem for R^− induced by (6) is min −tr(R^−) subject to the two trace constraints, which is unbounded below; the implemented update (18) is well-posed and solves a different problem. Consequently, the numerical experiments in Section 6 cannot be claimed to minimize (6). Either (6) must be amended to match the implemented update, or Algorithm 2 must be changed to solve the true subproblem, and the analysis redone.
  3. [§5.1, Lemma 2 and Appendix B] Lemma 2 asserts Lipschitz continuity of the solutions to (12) and (14), and the proof assumes unique minimizers. However, the projection Π_D onto the nonconvex complementarity set D={v+≤0,v−≤0,v+⊥v−} is not single-valued in general; for example, the point (−2,−2) has two projections, (0,−2) and (−2,0). No tie-breaking rule is specified, and the argument that A is full rank does not address the nonconvexity of D. Lemma 3 in Appendix C also assumes g(ℓ^s) is positive and bounded, but g can approach 0 on the boundary of the inequality constraint, making the log-barrier gradient unbounded. These gaps leave Theorem 1 unproved.
  4. [§5.2, Theorem 2] The proof of Theorem 2 invokes Tseng's BCD convergence framework and [48] for orthogonality constraints. But Algorithm 2 updates the L~_B^s block with a finite-horizon ADMM (Algorithm 1 with T iterations), which is an inexact block update; Tseng's framework requires exact block minimizers. The constraint ℓ+⊥ℓ− is a complementarity constraint, not the orthogonality constraint treated in [48]. Moreover, the boundedness of the level set used in the proof is false because of the unboundedness of (6). Thus convergence to a stationary point of (6) is not established.
minor comments (5)
  1. [§6.1.1] The definition of RelErr writes the numerator as ∥L̂−L*∥_2^2; it should presumably be the squared Frobenius norm, consistent with the denominator.
  2. [§6.2.1] After Fig. 2, the text says the learning problem 'becomes more tractable' as H increases; the reported degradation in F-score/RelErr suggests the opposite wording was intended.
  3. [Algorithm 2] The input lists the full signal matrix X and randomly selects the hidden nodes. In the stated problem only X_B is observed; the input specification and Step 2 should be consistent with the observation model.
  4. [§4.1, Eq. (8)] 'Vertical operator' appears to be a typo for 'orthogonality constraint' in the description of ℓ+⊥ℓ−.
  5. [§3.1] Assumptions AS2 and AS3 are stated for the full graph signals X, which are not observed. The subsequent formulation estimates the full total variation through correction terms; a sentence clarifying this modeling step would help the reader.

Circularity Check

0 steps flagged

No circularity: the derivation is self-contained; the unboundedness concern is a correctness issue, not circularity.

full rationale

The paper's derivation chain is a standard estimation formulation: it starts from a smoothness prior on signed graphs, introduces hidden-node correction terms via matrix block decompositions, imposes a column-sparsity prior on the correction matrices, and then solves a constrained optimization problem. The target quantities (the signed graph Laplacians on observed nodes) are the optimization variables, not quantities that have been pre-fitted or defined in terms of the outputs. Theorems 1 and 2 invoke external convergence frameworks (Tseng [44], Wang et al. [47], Fu et al. [48]) rather than relying on author-specific uniqueness or convergence claims. The self-citation to the authors' prior work [43] is used only for context and comparison in the introduction and Table 1; it is not load-bearing for the formulation, algorithm, or convergence proof. The skeptic's unboundedness observation about the term -tr(R^-) in problem (6), and the apparent inconsistency with the R^- update in (18), is a substantive correctness/well-posedness concern, but it is not a circular reduction: no output is equivalent to an input by construction, and no fitted parameter is renamed as a prediction. Therefore, by the circularity criteria defined for this analysis, there is no significant circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

No new physical entities are introduced. The auxiliary matrices P~^s and R^s are latent optimization variables, not new measurable quantities. The ledger shows a moderate number of hand-chosen regularization parameters and a set of domain assumptions, the most fragile being the sparse observed-hidden connectivity required for column-sparsity.

free parameters (5)
  • alpha_+, alpha_-
    Sparsity and density trade-off parameters in objective (6); no values or selection procedure reported in experiments.
  • sigma_+, sigma_-
    Column-sparsity (group lasso) regularization weights; no values reported.
  • eta_+, eta_-
    Log-barrier penalty parameters introduced to move the total-variation constraints into the objective; no values reported.
  • rho
    ADMM penalty parameter; no values reported.
  • number of hidden nodes H
    Input to Algorithm 2, treated as known. Real-world experiments set H=5; synthetic experiments vary H=1..5. Sensitivity to misspecification is not studied.
axioms (5)
  • domain assumption Signals are smooth over the positive-edge graph G+ and non-smooth over the negative-edge graph G- (AS2, AS3)
    Stated in Section 3.1; the total-variation objective is built directly on this assumption.
  • domain assumption Hidden node count is much smaller than observed node count (AS1: H << B)
    Needed to justify the low-rank structure of P^s and the whole block decomposition.
  • domain assumption The graph is sparse, so D~^s_B is low-rank and P~^s is column-sparse
    Invoked informally in Section 3.2 after Eq. (5); not formally stated as an assumption.
  • ad hoc to paper The convergence frameworks of Tseng [44] and Wang et al. [47] apply to the inexact BCD with non-convex complementarity constraints
    Used to prove Theorems 1 and 2, but the subproblem updates are not exact and the projection set D is non-convex; key conditions are asserted rather than verified.
  • domain assumption The sample covariance matrix C_B computed from observed signals equals the principal submatrix of the full covariance of the latent model
    The derivation requires C_B to be exactly the block of the true covariance; in practice it is only estimated from finite samples, introducing error not modeled in the problem.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Signed Graph Learning with Hidden Nodes." pith.science (2026). https://pith.science/paper/PVZZ6TPT

@misc{pith2026250909120,
  author       = {Pith},
  title        = {Pith review of: Signed Graph Learning with Hidden Nodes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PVZZ6TPT}},
  note         = {Machine review of arXiv:2509.09120}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Signed graphs, which are characterized by both positive and negative edge weights, have recently attracted significant attention in the field of graph signal processing (GSP). Existing works on signed graph learning typically assume that all graph nodes are available. However, in some specific applications, only a subset of nodes can be observed while the remaining nodes stay hidden. To address this challenge, we propose a novel method for identifying signed graph that accounts for hidden nodes, termed \textit{signed graph learning with hidden nodes under column-sparsity regularization} (SGL-HNCS). Our method is based on the assumption that graph signals are smooth over signed graphs, i.e., signal values of two nodes connected by positive (negative) edges are similar (dissimilar). Rooted in this prior assumption, the topology inference of a signed graph is formulated as a constrained optimization problem with column-sparsity regularization, where the goal is to reconstruct the signed graph Laplacian matrix without disregarding the influence of hidden nodes. We solve the constrained optimization problem using a tailored block coordinate descent (BCD) approach. Experimental results using synthetic data and real-world data demonstrate the efficiency of the proposed SGL-HNCS method.

Figures

Figures reproduced from arXiv: 2509.09120 by Hui Feng, Jian Wang, Rong Ye, Runhe Qiu, Xue-Qin Jiang.

Figure 1
Figure 1. Figure 1: An example of decomposing a signed graph into two unsigned graphs G + and G − in a social network. whose elements represent the weight of the edges. For an undirected graph, W is symmetric with zero diagonal entries, and the (i, j)-th entry Wi j is assigned a non-negative value if (i, j) ∈ E. Define the Laplacian matrix as L := D − W, where D is the diagonal degree matrix with entries Dii = PN j=1 Wi j bei… view at source ↗
Figure 2
Figure 2. Figure 2: Numerical validation of the proposed method under ER graph model with fixed M=50. (a) F-score of the recovered graphs for several methods as the number of hidden nodes increases. (b) RelErr of recovered graphs as the number of hidden nodes increases for different methods. 6.2.1. Number of Hidden Nodes In the first synthetic experiment, we present how the performance of different methods changes as the numb… view at source ↗
Figure 3
Figure 3. Figure 3: The designed signed graph based on ER model. 5 10 15 2 4 6 8 10 12 14 16 18 (a) Groundtruth L˜ + B 5 10 15 2 4 6 8 10 12 14 16 18 (b) SGL-HNLR Lˆ + B 5 10 15 2 4 6 8 10 12 14 16 18 (c) SGL-HNCS Lˆ + B 5 10 15 2 4 6 8 10 12 14 16 18 0 0.5 1 1.5 2 2.5 (d) scSGL Lˆ + B 5 10 15 2 4 6 8 10 12 14 16 18 (e) Groundtruth L˜ − B 5 10 15 2 4 6 8 10 12 14 16 18 (f) SGL-HNLR Lˆ − B 5 10 15 2 4 6 8 10 12 14 16 18 (g) SG… view at source ↗
Figure 4
Figure 4. Figure 4: The comparison of groundtruth and learned Laplacian matrices. The columns from the left to the right are the groundtruth Laplacians, the Laplacians learned by SGL-HNLR, the Laplacians learned by SGL-HNCS and the Laplacians learned by scSGL. The rows from the top to the bottom are the Laplacians for G + and G − , respectively. keeping the number of hidden nodes fixed at H = 2 [PITH_FULL_IMAGE:figures/full_… view at source ↗
Figure 5
Figure 5. Figure 5: Numerical validation of the proposed method under ER graph model with fixed H = 2. (a) F-score of the recovered graphs for several methods as the number of graph signals increases. (b) RelErr of the recovered graphs as the number of graph signals increases for different methods. 5 1 0 1 5 20 0. 00 0. 01 0. 02 0. 03 E r r o r Number of iteration M H = 1 H = 2 H = 5 (a) 5 1 0 1 5 20 0. 00 0. 01 0. 02 0. 03 E… view at source ↗
Figure 6
Figure 6. Figure 6: Results of convergence. (a) the Error metric for different numbers of hidden nodes H as iteration M increases. (b) the Error metric for different numbers of graph signals K as iteration M increases. 6.2.3. Convergence Finally, the convergence results of the proposed algorithm with respect to the variables, i.e., the number of hidden nodes H and the number of graph signals K, are shown in Fig. 6a and Fig. 6… view at source ↗
Figure 7
Figure 7. Figure 7: RelErr performance of the methods in learning signed graph from two real-world datasets. The left panel shows the performance for Wiki dataset and the right one shows the performance for Epinions dataset. 6.3. Results on Real-world Data This subsection presents the experimental results of our method on two real-world datasets. We use the Wiki dataset [50] and the Epinions dataset [51] to learn the signed g… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

52 extracted references

  1. [1]

    Campbell, C

    W. Campbell, C. Dagli, and C. Weinstein. Social network analysis with content and graphs. Lincoln Lab. J., 20(1):61–81, 2013

    2013

  2. [2]

    E. D. Kolaczyk.Statistical analysis of network data: Methods and models. Springer, New York, 2009

    2009

  3. [3]

    Liégeois, A

    R. Liégeois, A. Santos, V . Matta, D. Van De Ville, and A. H. Sayed. Revisiting correlation- based functional connectivity and its relationship with structural connectivity.Netw. Neu- rosci., 4(4):1235–1251, 2020

    2020

  4. [4]

    Namaki, A

    A. Namaki, A. Shirazi, R. Raei, and G. Jafari. Network analysis of a financial market based on genuine correlation and threshold method.Physica A., 390(21):3835–3841, 2011

    2011

  5. [5]

    D. I. Shuman, S. K. Narang, P. Frossard, A. Ortega, and P. Vandergheynst. The emerging field of signal processing on graphs: Extending high-dimensional data analysis to networks and other irregular domains.IEEE Signal Process. Mag., 30(3):83–98, 2013

    2013

  6. [6]

    Sandryhaila and J

    A. Sandryhaila and J. M. F. Moura. Discrete signal processing on graphs.IEEE Trans. Signal Process., 61(7):1644–1656, 2013

    2013

  7. [7]

    A. G. Marques, N. Kiyavash, J. M. F. Moura, D. Van De Ville, and R. Willett. Graph signal processing: Foundations and emerging directions [from the guest editors].IEEE Signal Process. Mag., 37(6):11–13, 2020

    2020

  8. [8]

    G. Leus, A. G. Marques, J. M. F Moura, A. Ortega, and D. I. Shuman. Graph signal process- ing: History, development, impact, and outlook.IEEE Signal Process. Mag., 40(4):49–60, 2023

    2023

  9. [9]

    Y . Yan, E. E. Kuruoglu, and M. A. Altinkaya. Adaptive sign algorithm for graph signal processing.Signal Process., 200:108662, 2022

    2022

  10. [10]

    W. Liu, H. Feng, F. Ji, and B. Hu. Online signed sampling of bandlimited graph signals. IEEE Trans. Signal Inf. Process. Netw., 10:131–146, 2024

    2024

  11. [11]

    J. H. Giraldo, S. Javed, M. Sultana, S. K. Jung, and T. Bouwmans. The emerging field of graph signal processing for moving object segmentation. InProc. Int. Workshop Frontiers Comput. Vis.(IW-FCV), pages 31–45, 2021

    2021

  12. [12]

    Dinesh, G

    C. Dinesh, G. Cheung, and I. V . Baji ´c. Point cloud denoising via feature graph Laplacian regularization.IEEE Trans. Image Process., 29:4143–4158, 2020

    2020

  13. [13]

    X. Song, L. Chai, and J. Zhang. Graph signal processing approach to QSAR/QSPR model learning of compounds.IEEE Trans. Pattern Anal. Mach. Intell., 44(4):1963–1973, 2020

    1963

  14. [14]

    J. Jin, J. Zhang, J. Tang, S. Liang, and Z. Qu. Spatio-temporal data mining with information integrity protection: Graph signal based air quality prediction. InProc. IEEE Int. Conf. Acoust., Speech Signal Process. (ICASSP), pages 5190–5194, 2024

    2024

  15. [15]

    J. Li, T. Wan, and W. Qiu. Time-varying sea surface temperature reconstruction leveraging low rank and joint smoothness constraints.J. Electron. Inf. Techn., 47(3):1–9, 2025. 22

    2025

  16. [16]

    Pavez and A

    E. Pavez and A. Ortega. Generalized Laplacian precision matrix estimation for graph signal processing. InProc. IEEE Int. Conf. Acoust., Speech Signal Process. (ICASSP), pages 6350–6354, 2016

    2016

  17. [17]

    G. B. Giannakis, Y . Shen, and G. V . Karanikolas. Topology identification and learning over graphs: Accounting for nonlinearities and dynamics.Proc. IEEE., 106(5):787–807, 2018

    2018

  18. [18]

    X. Dong, D. Thanou, M. Rabbat, and P. Frossard. Learning graphs from data: A signal representation perspective.IEEE Signal Process. Mag., 36(3):44–63, 2019

    2019

  19. [19]

    Mateos, S

    G. Mateos, S. Segarra, A. G. Marques, and A. Ribeiro. Connecting the dots: Identifying network structure via graph signal processing.IEEE Signal Process. Mag., 36(3):16–43, 2019

    2019

  20. [20]

    Z. Song, X. Yang, Z. Xu, and I. King. Graph-based semi-supervised learning: A compre- hensive review.IEEE Trans. Neural Netw. Learn. Syst., 34(11):8174–8194, 2022

    2022

  21. [21]

    H. E. Egilmez, E. Pavez, and A. Ortega. Graph learning from data under Laplacian and structural constraints.IEEE J. Sel. Topics Signal Process., 11(6):825–841, 2017

    2017

  22. [22]

    Kumar, J

    S. Kumar, J. Ying, J. V . de Miranda Cardoso, and D. P. Palomar. Structured graph learning via Laplacian spectral constraints.Adv. Neural. Inf. Process. Syst., 32:11647–11658, 2019

    2019

  23. [23]

    Javaheri, A

    A. Javaheri, A. Amini, F. Marvasti, and D. P. Palomar. Learning spatio-temporal graphical models from incomplete observations.IEEE Trans. Signal Process., 72:1361–1374, 2024

    2024

  24. [24]

    Kalofolias

    V . Kalofolias. How to learn a graph from smooth signals. InProc. Int. Conf. Artif. Intel. Statist. J. Mach. Learn. Res. (PMLR), pages 920–929, 2016

    2016

  25. [25]

    X. Dong, D. Thanou, P. Frossard, and P. Vandergheynst. Learning Laplacian matrix in smooth graph signal representations.IEEE Trans. Signal Process., 64(23):6160–6173, 2016

    2016

  26. [26]

    Bagheri, G

    S. Bagheri, G. Cheung, T. Eadie, and A. Ortega. Joint signal interpolation/time-varying graph estimation via smoothness and low-rank priors. InProc. IEEE Int. Conf. Acoust., Speech Signal Process. (ICASSP), pages 9646–9650, 2024

    2024

  27. [27]

    Segarra, A

    S. Segarra, A. G. Marques, G. Mateos, and A. Ribeiro. Network topology inference from spectral templates.IEEE Trans. Signal Inf. Process. Netw., 3(3):467–483, 2017

    2017

  28. [28]

    Dittrich and G

    T. Dittrich and G. Matz. Signal processing on signed graphs: Fundamentals and potentials. IEEE Signal Process. Mag., 37(6):86–98, 2020

    2020

  29. [29]

    Matz and T

    G. Matz and T. Dittrich. Learning signed graphs from data. InProc. IEEE Int. Conf. Acoust., Speech Signal Process. (ICASSP), pages 5570–5574, 2020

    2020

  30. [30]

    Girdhar and K

    N. Girdhar and K. K. Bharadwaj. Signed social networks: A survey. InInt. Conf. Adv. Comput. Data Sci. (ICACDS), pages 326–335, 2017

    2017

  31. [31]

    Karaaslanli, S

    A. Karaaslanli, S. Saha, S. Aviyente, and T. Maiti. scSGL: Kernelized signed graph learn- ing for single-cell gene regulatory network inference.Bioinformatics, 38(11):3011–3019, 2022. 23

    2022

  32. [32]

    Chandrasekaran, P

    V . Chandrasekaran, P. A. Parrilo, and A. S. Willsky. Latent variable graphical model selection via convex optimization. InAnnu. Allerton Conf. Commun., Control, Com- put.(Allerton), pages 1610–1613, 2010

    2010

  33. [33]

    Chang, T

    A. Chang, T. Yao, and G. I. Allen. Graphical models and dynamic latent factors for mod- eling functional brain connectivity. InProc. IEEE Data Sci. Wrksp. (DSW), pages 57–63, 2019

    2019

  34. [34]

    X. Yang, M. Sheng, Y . Yuan, and T. Q. S. Quek. Network topology inference from hetero- geneous incomplete graph signals.IEEE Trans. Signal Process., 69:314–327, 2021

    2021

  35. [35]

    Anandkumar, D

    A. Anandkumar, D. Hsu, A. Javanmard, and S. Kakade. Learning linear bayesian networks with latent variables. InProc. Int. Conf. Mach. Learn. (ICML), pages 249–257, 2013

    2013

  36. [36]

    Mei and M

    J. Mei and M. F. Moura. Silvar: Single index latent variable models.IEEE Trans. Signal Process., 66(11):2790–2803, 2018

    2018

  37. [37]

    Buciulea, S

    A. Buciulea, S. Rey, C. Cabrera, and A. G. Marques. Network reconstruction from graph- stationary signals with hidden variables. InAsilomar Conf. Signals, Syst., Comput. (AC- SSC), pages 56–60, 2019

    2019

  38. [38]

    Buciulea, S

    A. Buciulea, S. Rey, and A. G. Marques. Learning graphs from smooth and graph-stationary signals with hidden variables.IEEE Trans. Signal Inf. Process. Netw., 8:273–287, 2022

    2022

  39. [39]

    J. H. Giraldo, A. Mahmood, B. Garcia-Garcia, D. Thanou, and T. Bouwmans. Recon- struction of time-varying graph signals via Sobolev smoothness.IEEE Trans. Signal Inf. Process. Netw., 8:201–214, 2022

    2022

  40. [40]

    A. G. Marques, S. Segarra, G. Leus, and A. Ribeiro. Stationary graph processes and spectral estimation.IEEE Trans. Signal Process., 65(22):5911–5926, 2017

    2017

  41. [41]

    Perraudin and P

    N. Perraudin and P. Vandergheynst. Stationary signal processing on graphs.IEEE Trans. Signal Process., 65(13):3462–3477, 2017

    2017

  42. [42]

    S. Rey, A. Buciulea, M. Navarro, S. Segarra, and A. G. Marques. Joint inference of mul- tiple graphs with hidden variables from stationary graph signals. InProc. IEEE Int. Conf. Acoust., Speech Signal Process. (ICASSP), pages 5817–5821, 2022

    2022

  43. [43]

    R. Ye, X. Q. Jiang, H. Feng, J. Wang, R. Qiu, and X. Hou. Time-varying graph learning from smooth and stationary graph signals with hidden nodes.EURASIP J. Adv. Signal Process., 2024(1):33, 2024

    2024

  44. [44]

    P. Tseng. Convergence of a block coordinate descent method for nondifferentiable mini- mization.J. Optim. Theory Appl., 103(9):475–494, 2001

    2001

  45. [45]

    Grant and S

    M. Grant and S. Boyd. CVX: Matlab software for disciplined convex programming, version 2.1 beta. http://cvxr.com/cvx, 2013

    2013

  46. [46]

    Scheel and S

    H. Scheel and S. Scholtes. Mathematical programs with complementarity constraints: Sta- tionarity, optimality, and sensitivity.Math. Oper. Res., 25(1):1–22, 2000. 24

    2000

  47. [47]

    Y . Wang, W. Yin, and J. Zeng. Global convergence of ADMM in nonconvex nonsmooth optimization.J. Sci. Comput., 78:29–63, 2019

    2019

  48. [48]

    X. Fu, K. Huang, M. Hong, N. D. Sidiropoulos, and A. M. C. So. Scalable and flex- ible multiview MAX-V AR canonical correlation analysis.IEEE Trans. Signal Process., 65(16):4150–4165, 2017

    2017

  49. [49]

    Erd6s and A

    P. Erd6s and A. Rényi. On the evolution of random graphs.Publ. Math. Inst. Hungar. Acad. Sci., 5(1):17–61, 1960

    1960

  50. [50]

    Leskovec, D

    J. Leskovec, D. Huttenlocher, and J. Kleinberg. Governance in social media: A case study of the Wikipedia promotion process. InProc. Int. AAAI Conf. Web Soc. Media. (ICWSM), pages 98–105, 2010

    2010

  51. [51]

    Massa and P

    P. Massa and P. Avesani. Controversial users demand local trust metrics: An experimental study on epinions. com community. InProc. Int. AAAI Conf. Web Soc. Media. (ICWSM), pages 121–126, 2005

    2005

  52. [52]

    M. Hong, M. Razaviyayn, Z. Luo, and J. Pang. A unified algorithmic framework for block- structured optimization involving big data: With applications in machine learning and sig- nal processing.IEEE Signal Process. Mag., 33(1):57–77, 2015. 25

    2015

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.