Pith. sign in

REVIEW 4 major objections 5 minor 24 references

Signed-Graph Recommendation as Structural Consistency Maximization

T0 review · 4 major / 5 minor · reviewed 2026-07-08 · grok-4.5

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

Pith's one-line read Signed social recommendation can be cast as maximizing consistency across structure, propagation, and semantics, and SSC-Loop enforces that objective.

desk verdict Clean packaging of signed social rec as multi-layer consistency maximization; useful if the Epinions ablations hold, but the 'fundamental inconsistency' diagnosis is asserted harder than the abstract can support. read the letter →

arxiv 2607.05952 v1 pith:SXCI3Q5C submitted 2026-07-07 cs.SI cs.AI

Zifan Wang , Siyu Chen , Wenzhuo Song This is my paper
classification cs.SIcs.AI
keywords signedsocialrecommendationstructuralconsistencygraphstrustanddistrustgraphneuralnetworkscontrastivelearningdatasparsity
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

Existing signed social recommenders that model both trust and distrust are limited by structural noise and data sparsity. The paper identifies a fundamental inconsistency that appears across three layers at once: the observed signed graph, the way messages are propagated over it, and the semantic embeddings that are ultimately used for rating prediction. That inconsistency, the authors argue, is what produces biased representations. They therefore reframe the entire task as maximization of structural consistency and introduce SSC-Loop, a single loop that restores agreement among the three layers. ESA-DA adaptively refines the noisy topology, a P/N/O propagation scheme separately handles positive, negative, and observed relations, and a contrastive objective pulls the resulting embeddings toward reliable social semantics. On Epinions the framework improves explicit signed social rating prediction; auxiliary results on Slashdot under a derived link-existence setting further indicate that the same machinery can exploit signed structure. If the diagnosis holds, consistency maximization becomes a unifying design principle for signed social recommendation under sparsity and noise.

What carries the argument

SSC-Loop, a closed three-module loop that turns recommendation into iterative consistency maximization: ESA-DA (edge-structure-aware dual adaptation) cleans and adapts the signed graph, P/N/O propagation separately routes positive, negative, and observed signals, and a contrastive objective enforces semantic agreement between the refined structure and the learned embeddings.

What would settle it

An ablation that freezes the observed signed graph, replaces P/N/O with ordinary signed propagation, and drops the contrastive semantic term should erase most of the reported Epinions rating-prediction gains under identical capacity and training budget; if the gains largely remain, the three-layer consistency diagnosis is not necessary.

Watch

Extended reading notes

Core claim

The performance of signed social recommendation is limited by inconsistency among the observed signed topology, the propagation rules applied to it, and the learned semantic embeddings; maximizing consistency across these three layers with ESA-DA structure refinement, P/N/O signed propagation, and contrastive semantic alignment produces more reliable representations and stronger rating predictions under sparsity and noise.

Load-bearing premise

The premise that a fundamental inconsistency across structural, propagation, and semantic layers is the primary driver of biased representations under sparsity and noise, and that jointly maximizing consistency is the right corrective.

Editorial extensions

If this is right

  • Signed social rating prediction improves when structure refinement, signed message passing, and semantic alignment are optimized jointly rather than in isolation.
  • Models that treat the observed signed graph as fixed remain biased under sparsity; adaptive topology refinement becomes a necessary first stage.
  • The same consistency objective transfers from rating prediction to signed link-existence tasks, as suggested by the Slashdot results.
  • Future signed recommenders can be diagnosed by measuring residual inconsistency across the three layers and intervening only where disagreement is largest.

Reading between the lines

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

  • The three-layer consistency view may extend to any multi-relational GNN setting in which topology, propagation, and labels can disagree, not only signed social graphs.
  • If ESA-DA accounts for most of the lift, simpler adaptive edge reweighting or sparsification baselines may capture a large fraction of the gain without the full loop.
  • A practical diagnostic would compute layer-wise inconsistency scores on any signed recommender to decide whether the next intervention should target structure, propagation, or semantics.
  • The contrastive semantic term could be swapped for other alignment losses without abandoning the overall consistency-maximization thesis.
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 argues that signed social recommendation is hindered by a multi-layer inconsistency among structural topology, signed propagation, and learned semantics under sparsity and noise. It casts the task as structural consistency maximization and proposes SSC-Loop with three modules: ESA-DA (edge selection and dual adaptation of the signed graph), a P/N/O multi-channel propagation scheme (positive, negative, and other/unsigned paths), and a contrastive objective for semantic consistency. On Epinions the method is evaluated for explicit signed social rating prediction against several baselines; auxiliary Slashdot experiments use a derived link-existence setup. The authors report strong Epinions results and claim that jointly closing the three consistency gaps yields more reliable representations than treating the observed signed graph as fixed.

Significance. If the diagnosis and causal attribution hold, the work offers a coherent organizing principle for signed social recommendation and a concrete three-module architecture that unifies graph editing, signed message passing, and contrastive regularization. Strengths include an explicit multi-layer framing, public code, and evaluation on a standard signed rating dataset (Epinions) plus an auxiliary signed-structure setting (Slashdot). The contribution is primarily methodological and empirical rather than theoretical; its lasting value depends on whether the consistency framing is necessary (vs. generic capacity/regularization gains) and whether the reported gains survive careful ablations and matched baselines.

major comments (4)
  1. The central claim that a 'fundamental inconsistency' across structural, propagation, and semantic layers is the primary driver of biased representations is asserted as motivation but is not isolated from ordinary alternatives (label noise, cold-start, under-tuned baselines, evaluation leakage). The manuscript needs a direct diagnostic: e.g., a quantitative inconsistency measure that tracks prediction error before/after each module, or a controlled comparison showing that matched capacity without the consistency objectives fails where SSC-Loop succeeds. Without this, the unified framing risks overclaiming relative to independent engineering gains from graph editing, multi-channel GNN capacity, and contrastive regularization.
  2. Causal attribution of Epinions gains to jointly maximizing the three consistencies is not yet load-bearing. Full ablations that remove ESA-DA, P/N/O, and the contrastive term one at a time (and in combination), plus a capacity-matched baseline that adds generic graph editing or multi-view contrast without the signed-consistency objectives, are required. Report effect sizes with error bars or multiple random seeds; if gains remain after matching capacity and regularization, the consistency story is supported; if not, the framing should be softened.
  3. Slashdot is evaluated under a derived link-existence setting rather than explicit signed rating prediction. The abstract and claims should clearly separate this auxiliary protocol from the primary Epinions rating task so that 'strong performance on signed social recommendation' is not overstated. Specify how positive/negative labels and non-edges are constructed, whether structure used at train time can leak into the derived labels, and report the same metrics/baselines as on Epinions where applicable.
  4. ESA-DA edits the observed signed graph; any evaluation that reuses edited edges or that scores consistency with metrics aligned to the same edit objective can become self-confirming. Clarify the train/eval split relative to structure adaptation (is the test graph frozen? are only training edges adapted?), and ensure that rating metrics are not confounded by structure-edit leakage. If structure is adapted using rating labels, discuss the risk of label leakage into topology.
minor comments (5)
  1. Hyperparameters (loss weights, ESA-DA thresholds, contrastive temperature, propagation depth) should be listed in a single table with search ranges and selected values for reproducibility.
  2. Notation for P/N/O channels and the precise form of the contrastive pairs (what constitutes a positive vs. negative view under signed edges) should be stated with equations in one place rather than distributed across sections.
  3. Baseline implementations and tuning protocol should be documented (same search budget as SSC-Loop) to reduce the risk that gains come from under-tuned competitors.
  4. Figures/tables should report standard deviations over seeds; single-run point estimates are insufficient for the strength of claim made on Epinions.
  5. Related work should more explicitly position ESA-DA against prior signed-graph denoising / edge-sign prediction and P/N/O against existing multi-channel signed GNNs, to clarify novelty boundaries.

Simulated Author's Rebuttal

4 responses · 0 unresolved

We thank the referee for a careful and constructive review. The comments correctly press on causal attribution, evaluation protocol clarity, and the risk of overclaiming the multi-layer consistency framing. We agree that several of these points require stronger diagnostics, fuller ablations with capacity-matched controls, and clearer separation of the Epinions rating task from the auxiliary Slashdot protocol. We will revise the manuscript accordingly: adding quantitative consistency diagnostics, completing module-wise and capacity-matched ablations with multi-seed statistics, clarifying Slashdot construction and claims, and making the ESA-DA train/eval protocol and leakage discussion explicit. Below we respond point by point.

read point-by-point responses

  1. Referee: The central claim that a 'fundamental inconsistency' across structural, propagation, and semantic layers is the primary driver of biased representations is asserted as motivation but is not isolated from ordinary alternatives (label noise, cold-start, under-tuned baselines, evaluation leakage). The manuscript needs a direct diagnostic: e.g., a quantitative inconsistency measure that tracks prediction error before/after each module, or a controlled comparison showing that matched capacity without the consistency objectives fails where SSC-Loop succeeds. Without this, the unified framing risks overclaiming relative to independent engineering gains from graph editing, multi-channel GNN capacity, and contrastive regularization.

    Authors: We agree that the current draft motivates the multi-layer inconsistency primarily through qualitative argument and end-to-end gains, which is insufficient to isolate it from ordinary alternatives (noise, cold-start, capacity, or under-tuned baselines). The framing is intended as an organizing principle for the three modules, not as a fully established causal law; we will soften language that presents it as already proven. In revision we will add a direct diagnostic section: (i) a quantitative structural/propagation/semantic inconsistency measure computed before and after each module and correlated with rating error; (ii) a controlled comparison against capacity-matched variants that retain graph editing, multi-channel message passing, or contrastive regularization without the signed-consistency objectives. Where the diagnostic supports the framing we will state that clearly; where gains are partly attributable to generic capacity/regularization we will narrow the claim. This addresses the overclaiming risk without abandoning the architectural motivation. revision: yes

  2. Referee: Causal attribution of Epinions gains to jointly maximizing the three consistencies is not yet load-bearing. Full ablations that remove ESA-DA, P/N/O, and the contrastive term one at a time (and in combination), plus a capacity-matched baseline that adds generic graph editing or multi-view contrast without the signed-consistency objectives, are required. Report effect sizes with error bars or multiple random seeds; if gains remain after matching capacity and regularization, the consistency story is supported; if not, the framing should be softened.

    Authors: This is a fair and load-bearing request. The current ablations are incomplete for causal attribution: they do not fully remove ESA-DA, P/N/O, and the contrastive term both singly and in combination, nor do they include capacity-matched generic editing / multi-view contrast controls, and multi-seed error bars are not reported at the level the referee asks for. We will add: (1) leave-one-module-out and leave-two-out ablations for ESA-DA, P/N/O, and the contrastive objective; (2) capacity-matched baselines that add generic graph editing or multi-view contrast without signed-consistency targets; (3) mean ± std over multiple random seeds and effect-size reporting on Epinions. If residual gains remain after matching capacity and regularization, we will treat that as support for the joint-consistency story; if not, we will soften the causal language in the abstract, introduction, and conclusion to methodological contribution plus empirical gains under the proposed architecture. We will not claim joint maximization as established unless the new tables support it. revision: yes

  3. Referee: Slashdot is evaluated under a derived link-existence setting rather than explicit signed rating prediction. The abstract and claims should clearly separate this auxiliary protocol from the primary Epinions rating task so that 'strong performance on signed social recommendation' is not overstated. Specify how positive/negative labels and non-edges are constructed, whether structure used at train time can leak into the derived labels, and report the same metrics/baselines as on Epinions where applicable.

    Authors: We agree. Slashdot is an auxiliary signed-structure / derived link-existence protocol, not explicit signed social rating prediction; the abstract already flags this, but the body and claim language still risk conflating the two. We will: (i) keep Epinions as the sole primary rating-prediction evaluation and explicitly label Slashdot as auxiliary throughout abstract, intro, experiments, and conclusion; (ii) detail construction of positive/negative labels and non-edges, the train/test split relative to observed structure, and any risk that structure used at train time leaks into derived labels; (iii) report metrics and baselines that are applicable under the link-existence setup, and avoid implying parity with Epinions rating metrics where the tasks differ. We will not use Slashdot to underwrite the phrase 'strong performance on signed social recommendation' without the Epinions qualifier. revision: yes

  4. Referee: ESA-DA edits the observed signed graph; any evaluation that reuses edited edges or that scores consistency with metrics aligned to the same edit objective can become self-confirming. Clarify the train/eval split relative to structure adaptation (is the test graph frozen? are only training edges adapted?), and ensure that rating metrics are not confounded by structure-edit leakage. If structure is adapted using rating labels, discuss the risk of label leakage into topology.

    Authors: We agree that ESA-DA requires an explicit leakage protocol. In revision we will state clearly: structure adaptation is performed using only training interactions/edges; the test graph (and test ratings) are frozen and never used to select or reweight edges for evaluation; rating metrics are computed solely on held-out user–item ratings, not on consistency scores aligned to the edit objective. We will also discuss the residual risk that adapting topology with training rating labels can inject label information into the graph used at train time, and describe the safeguards we use (train-only adaptation, frozen test structure, no reuse of edited edges as evaluation targets). If any experiment currently blurs this boundary, we will correct the protocol and re-report the affected numbers. This removes the self-confirming evaluation concern as far as the rating task is concerned. revision: yes

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; consistency-maximization framing is evaluated on independent rating-prediction metrics, not on the consistency objective itself.

full rationale

From the abstract and stated claims, SSC-Loop casts signed social recommendation as maximization of structural consistency via three modules (ESA-DA, P/N/O propagation, contrastive semantic consistency). The load-bearing empirical claim is strong performance on explicit signed social rating prediction on Epinions (plus auxiliary derived link-existence on Slashdot). That evaluation target is external to the consistency objectives: rating prediction accuracy is not the same quantity as the structural/propagation/semantic consistency losses being optimized, so the reported gains are not forced by construction from the objective definitions. There is no fitted scalar or ratio redefined as a prediction; no uniqueness theorem imported from the authors; no load-bearing self-citation chain visible; and no equation-level self-definition where X is defined as Y and then Y is claimed to derive X. Packaging graph editing, signed multi-channel propagation, and contrastive learning under a unifying consistency narrative is ordinary methodological framing, not renaming of a known empirical law into a first-principles derivation. The asserted multi-layer inconsistency as primary cause of bias is a soft modeling assumption (correctness risk), not circularity. Absent equation-level reductions of outputs to inputs, the honest finding is no significant circularity.

Assumptions & free parameters 1 free parameters · 3 assumptions · 2 invented entities

Abstract-only audit. Free parameters (loss weights, thresholds in ESA-DA, temperature in contrastive loss, propagation depths) are almost certainly present in any such GNN pipeline but not named here. Core domain assumptions: signed edges encode usable trust/distrust signal; observed social graphs are noisy but partially informative; consistency across structure/propagation/semantics is a valid training target for rating prediction. No new physical entities; invented pieces are algorithmic modules, not particles or forces. Full ledger would expand after reading methods and hyperparameter sections.

free parameters (1)
  • Unspecified SSC-Loop hyperparameters (loss weights, ESA-DA thresholds, contrastive temperature, propagation depth)
    Any multi-module GNN+contrastive+structure-learning system fits several scalars; abstract does not list them or claim they are fixed a priori. Treated as free until full text shows otherwise.
assumptions (3)
  • domain assumption Observed signed social graphs contain recoverable trust/distrust signal despite structural noise and sparsity.
    Stated as the setting of the work; without it, consistency maximization has no useful target.
  • ad hoc to paper A fundamental inconsistency across structural, propagation, and semantic layers causes biased representations in existing signed social recommenders.
    Abstract presents this diagnosis as the motivation for SSC-Loop; it is a paper-specific claim, not a standard theorem.
  • ad hoc to paper Maximizing structural consistency (via structure adaptation, P/N/O propagation, and contrastive semantics) improves signed rating prediction.
    This is the design hypothesis of the framework; empirical support is asserted but not inspectable from the abstract alone.
invented entities (2)
  • ESA-DA module
    purpose: Enforce structural consistency by adapting/denoising the signed graph rather than treating topology as fixed.
    Named algorithmic component introduced by the paper; independent evidence would be ablations and external benchmarks, not available in abstract.
  • P/N/O propagation mechanism
    purpose: Separate positive, negative, and other channels to achieve propagation consistency on signed graphs.
    Paper-specific propagation design; may extend known signed GNN channels but is presented as a dedicated module of SSC-Loop.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Signed-Graph Recommendation as Structural Consistency Maximization." pith.science (2026). https://pith.science/paper/SXCI3Q5C

@misc{pith2026260705952,
  author       = {Pith},
  title        = {Pith review of: Signed-Graph Recommendation as Structural Consistency Maximization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SXCI3Q5C}},
  note         = {Machine review of arXiv:2607.05952}
}
read the original abstract

While signed social recommendation has shown great potential by modeling both trust and distrust relations, its effectiveness is often hindered by structural noise and data sparsity. In this work, we first identify a fundamental inconsistency across the structural, propagation, and semantic layers of existing models, which leads to biased representations learned from sparse or noisy datasets. Furthermore, we observe that most existing methods treat the observed graph as fixed, failing to bridge the gap between noisy topologies and reliable social semantics. To address these issues, we propose a unified framework named SSC-Loop that treats signed social recommendation as the maximization of structural consistency. SSC-Loop includes three dedicated modules: ESA-DA for structural consistency, a P/N/O propagation mechanism for propagation consistency, and a contrastive learning objective for semantic consistency. Experiments on Epinions demonstrate that SSC-Loop achieves strong performance on explicit signed social rating prediction, while auxiliary results on Slashdot under a derived link-existence setting further suggest its ability to exploit signed social structures. Source code is available at https://github.com/Refrainwww/SSC-Loop.

Figures

Figures reproduced from arXiv: 2607.05952 by the authors.

Figure 1
Figure 1. The overall architecture of SSC-Loop. cv|u = X i∈I(u) α (v) ui zi . (2) If I(u) is empty, we set cv|u = ev. The edge compatibility score is then: ruv = cos(eu, cv|u). (3) We map this score into polarity-aware scores via a temperature-scaled sigmoid: p + uv = σ ruv τ  , p− uv = σ  − ruv τ  , (4) where τ denotes the temperature parameter. A larger p + uv indicates stronger evidence for a positive relation, while a… view at source ↗
Figure 2
Figure 2. Detailed analysis of SSC-Loop. On Slashdot, SSC-Loop obtains the best RMSE and the second-best MAE under the derived binary interaction setting. Since Slashdot does not contain real item ratings, these results are used only as auxiliary evidence that SSC-Loop can exploit signed social structures beyond explicit rating prediction. 4.4 Ablation Study and Further Analysis of SSC-Loop [PITH_FULL_IMAGE:figures/full_fig_… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

24 extracted references · 24 canonical work pages

  1. [1]

    Social Network Anal- ysis and Mining3(4), 1113–1133 (2013)

    Tang, J., Hu, X., Liu, H.: Social recommendation: a review. Social Network Anal- ysis and Mining3(4), 1113–1133 (2013)

    work page 2013

  2. [2]

    In: Proceedings of the Fourth ACM International Conference on Web Search and Data Mining, pp

    Ma, H., Zhou, D., Liu, C., Lyu, M.R., King, I.: Recommender systems with social regularization. In: Proceedings of the Fourth ACM International Conference on Web Search and Data Mining, pp. 287–296. ACM, New York (2011)

    work page 2011

  3. [3]

    In: The World Wide Web Conference, pp

    Fan, W., Ma, Y., Li, Q., He, Y., Zhao, E., Tang, J., Yin, D.: Graph neural networks for social recommendation. In: The World Wide Web Conference, pp. 417–426. ACM, New York (2019)

    work page 2019

  4. [4]

    In: Proceedings of the 19th International Conference on World Wide Web, pp

    Leskovec, J., Huttenlocher, D., Kleinberg, J.: Predicting positive and negative links in online social networks. In: Proceedings of the 19th International Conference on World Wide Web, pp. 641–650. ACM, New York (2010)

    work page 2010

  5. [5]

    In: Proceedings of the 2017 SIAM International Conference on Data Mining, pp

    Wang, S., Tang, J., Aggarwal, C., Chang, Y., Liu, H.: Signed network embedding in social media. In: Proceedings of the 2017 SIAM International Conference on Data Mining, pp. 327–335. SIAM (2017)

    work page 2017

  6. [6]

    In: International Conference on Artificial Neural Networks, pp

    Huang, J., Shen, H., Hou, L., Cheng, X.: Signed graph attention networks. In: International Conference on Artificial Neural Networks, pp. 566–577. Springer, Cham (2019)

    work page 2019

  7. [7]

    In: Proceedings of the 2007 ACM Conference on Recommender Systems, pp

    Massa, P., Avesani, P.: Trust-aware recommender systems. In: Proceedings of the 2007 ACM Conference on Recommender Systems, pp. 17–24. ACM, New York (2007)

    work page 2007

  8. [8]

    The Journal of Psychology21(1), 107–112 (1946)

    Heider, F.: Attitudes and cognitive organization. The Journal of Psychology21(1), 107–112 (1946)

    work page 1946

Show all 24 references
  1. [9]

    Jamali, M., Ester, M.: A matrix factorization technique with trust propagation for recommendationinsocialnetworks.In:ProceedingsoftheFourthACMConference on Recommender Systems, pp. 135–142. ACM, New York (2010)

    2010

  2. [10]

    arXiv preprint arXiv:2604.12735 (2026)

    Wang, Z., Yu, Z., Zhu, Y., Zhao, B., Liang, H., Wang, T., Xia, W., Zhang, J., Liu, Z., Ma, H., Ma, F., Tian, Q.: AffectAgent: Collaborative multi-agent rea- soning for retrieval-augmented multimodal emotion recognition. arXiv preprint arXiv:2604.12735 (2026)

    2026 arXiv

  3. [11]

    arXiv preprint arXiv:2605.18884 (2026) Signed-Graph Recommendation as Structural Consistency Maximization 13

    Wang, Z., Zhao, B., Zhu, Y., Liu, Z., Ma, H., Zhang, R., Ding, S., Xie, Q., Yu, Z.: Navigating the emotion tree: Hierarchical hyperbolic RAG for multimodal emotion recognition. arXiv preprint arXiv:2605.18884 (2026) Signed-Graph Recommendation as Structural Consistency Maximization 13

    2026 arXiv

  4. [12]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Zhao, B., Guo, D., Cao, J., Xu, Y., Zou, B., Tan, T., Sun, Y., Yu, Z.: Phase-net: Physics-grounded harmonic attention system for efficient remote photoplethysmog- raphy measurement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 2119...

    2026

  5. [13]

    arXiv preprint arXiv:2605.19950 (2026)

    Zhao, B., Ye, F., Ji, Y., Zhao, S., Peng, X., Yu, Z.: AffectVerse: Emotional world models for multimodal affective computing. arXiv preprint arXiv:2605.19950 (2026)

    2026 arXiv

  6. [14]

    Yang,B.,Lei,Y.,Liu,J.,Li,W.:Socialcollaborativefilteringbytrust.IEEETrans- actions on Pattern Analysis and Machine Intelligence39(8), 1633–1647 (2017)

    2017

  7. [15]

    In: Second European Network Intelligence Confer- ence, pp

    Bai, T., Yang, B., Li, F.: TDRec: Enhancing social recommendation using both trust and distrust information. In: Second European Network Intelligence Confer- ence, pp. 60–66. IEEE (2015)

    2015

  8. [16]

    In: Proceedings of the 25th International Conference on World Wide Web, pp

    Tang, J., Aggarwal, C., Liu, H.: Recommendations in signed social networks. In: Proceedings of the 25th International Conference on World Wide Web, pp. 31–40. International World Wide Web Conferences Steering Committee (2016)

    2016

  9. [17]

    In: Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval, pp

    He, X., Deng, K., Wang, X., Li, Y., Zhang, Y., Wang, M.: LightGCN: Simplifying and powering graph convolution network for recommendation. In: Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval, pp. 639–648. ACM, New ...

    2020

  10. [18]

    In: Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pp

    Chen, S., Chen, J., Zhou, S., Wang, B., Han, S., Su, C., Yuan, Y., Wang, C.: SIGformer: Sign-aware graph transformer for recommendation. In: Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pp. 1274–1284. ACM, New...

    2024

  11. [19]

    In: Jin, C., Yang, S., Shang, X., Wang, H., Zhang, Y

    Zeng, X., Chang, C., Tang, F., Wu, Z., Tang, Y.: Popularity-aware graph neural network with global context for session-based recommendation. In: Jin, C., Yang, S., Shang, X., Wang, H., Zhang, Y. (eds.) Web Information Systems and Applica- tions – WISA 2024. LNCS, vol. 14883, p...

    2024

  12. [20]

    Neurocomputing319, 42–54 (2018)

    Song, W., Wang, S., Yang, B., Lu, Y., Zhao, X., Liu, X.: Learning node and edge embeddings for signed networks. Neurocomputing319, 42–54 (2018)

    2018

  13. [21]

    Neurocomputing421, 329–339 (2021)

    Song, W., Chen, H., Liu, X., Jiang, H., Wang, S.: Hyperbolic node embedding for signed networks. Neurocomputing421, 329–339 (2021)

    2021

  14. [22]

    In: Knowledge Science, Engineering and Management, pp

    Zhao, X., Chen, H., Liu, X., Tan, X., Song, W.: Block modelling and learning for structure analysis of networks with positive and negative links. In: Knowledge Science, Engineering and Management, pp. 396–402. Springer (2018)

    2018

  15. [23]

    Knowledge-Based Systems195, 105714 (2020)

    Liu, X., Song, W., Musial, K., Zhao, X., Zuo, W., Yang, B.: Semi-supervised stochastic blockmodel for structure analysis of signed networks. Knowledge-Based Systems195, 105714 (2020)

    2020

  16. [24]

    ACM Transactions on Knowledge Discov- ery from Data19(3), 55:1–55:35 (2025)

    Liu, X., Song, W., Musial, K., Li, Y., Zhao, X., Yang, B.: Stochastic block models for complex network analysis: A survey. ACM Transactions on Knowledge Discov- ery from Data19(3), 55:1–55:35 (2025)

    2025

Pith tools

Reviewed July 8, 2026 · model on record in the stance chip above.