Neighborhood Overlap-Aware High-Order Graph Neural Network for Dynamic Graph Learning

Ling Wang

REVIEW 4 major objections 4 minor 68 references

Neighborhood Overlap-Aware High-Order Graph Neural Network for Dynamic Graph Learning

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

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

Pith's one-line read A dynamic graph GNN that scores neighborhood overlap and folds it into high-order message passing beats six baselines on link prediction.

desk verdict A plausible but unvalidated overlap-inspired attention GNN; the reported gains don't support the mechanism until ablations and validation are added. read the letter →

arxiv 2506.06728 v1 pith:A5MVBBD7 submitted 2025-06-07 cs.SI

Ling Wang This is my paper

classification cs.SI

keywords dynamicgraphlearninglinkpredictionneighborhoodoverlaphigh-orderneuralnetworktensorproductmessagepassingstructuralfeatures

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper's goal is to establish that dynamic graph learning should treat neighborhood overlap as a first-class signal, not just edge existence or feature similarity. It introduces NO-HGNN, which learns per-node structural features, scores each node pair by their dot product as an overlap correlation, and uses those scores to weight message passing inside a tensor-product high-order GNN. The paper reports that this mechanism improves link prediction over six static and dynamic baselines on two real-world dynamic graphs, with consistent gains on both. If the claim holds, any dynamic GNN that ignores overlap is leaving usable structural information on the table.

What carries the argument

The load-bearing object is the neighborhood-overlap correlation score $p_{ijt} = o_{it} \cdot o_{jt}$, where $o_{it}$ is a learned structural-feature vector for node $i$ at time $t$. The score is computed over structural features extracted from the multi-hop adjacency tensor $B = \sum_{k=1}^{K} A^k$; it is then softmax-normalized across neighbors and used as the aggregation-weight tensor $P$ inside a tensor-product HGNN. In other words, the overlap score converts structural overlap into message-passing weights, which is the mechanism the paper credits for the accuracy gains.

What would settle it

On the two datasets, compute exact shared-neighborhood counts from $B = \sum_{k=1}^{K} A^k$ for all node pairs per timestamp and compare them with the model's learned scores $p_{ijt}$; if the rank correlation is near zero while link prediction gains persist, the stated overlap mechanism is not the cause. A second check is to randomize or remove the structural feature generator and retrain: if accuracy does not fall, the claimed component is not load-bearing.

Watch

Extended reading notes

Core claim

Stated on the paper's own terms, the central claim is that the overlap between two nodes' neighborhoods, captured by $p_{ijt} = o_{it} \cdot o_{jt}$, is a signal dynamic GNNs currently underuse, and that injecting this signal into message passing improves link prediction. The model first forms a multi-hop overlap tensor $B = \sum_{k=1}^K A^k$, learns node-level structural features $O = [o_{it}]$ from $B$ with an MLP generator, computes a pairwise overlap-score tensor, softmax-normalizes it into aggregation weights $P$, and feeds $P$ into a tensor-product HGNN as the message-passing operator. The reported evidence is that this model outperforms all six baselines on both datasets: on D1, F1 and Accuracy rise from 0.8247 and 0.8002 for the strongest dynamic baseline to 0.8301 and 0.8125, and on D2 from 0.8134 and 0.7935 to 0.8266 and 0.8094.

Load-bearing premise

The load-bearing premise is that the dot product of learned structural features actually measures neighborhood overlap; the paper does not verify this, so if the learned features do not encode shared neighborhoods, NO-HGNN reduces to a generic softmax-attention GNN and the claimed reason for the accuracy gains is unsupported.

Editorial extensions

If this is right

If the reported gains are real, dynamic GNNs that weight messages by edge existence or raw feature correlation are leaving accuracy on the table; overlap-aware weighting is a drop-in modification to the aggregation step.
The overlap score is computed from the adjacency tensor alone, so the method requires no extra supervision beyond the link labels used for training.
The consistent gains across two datasets with different sizes and edge densities suggest the mechanism is not an artifact of one graph.
The architecture demonstrates that structural priors such as overlap can be embedded in high-order message passing without changing the downstream link prediction objective.

Reading between the lines

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

Because the paper does not run an ablation that replaces learned structural features with exact overlap counts, whether the learned generator is essential remains open; a natural follow-up is to swap $o_{it}$ for fixed one-hot or raw-overlap features and measure the drop.
The score is a dot product followed by softmax, which is structurally similar to attention; if the learned vectors do not track true overlap, the model could be succeeding as a generic attention GNN. A rank-correlation test between $p_{ijt}$ and exact shared-neighbor counts would distinguish these explanations.
The same overlap-aware weighting could be tested on neighboring tasks such as temporal edge classification or anomaly detection, but the paper only provides evidence for link prediction.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

The one-line take: this is a modest, coherent combination of known pieces, but the paper's central claim—that its learned score actually measures neighborhood overlap—is unsupported, and the experiments can't carry that weight.

What's new: NO-HGNN takes a tensor-product HGNN backbone and adds a softmax-normalized dot product of learned structural features as aggregation weights. Equations 3–4 sum adjacency powers to capture multi-hop neighborhoods, and Eqs. 6–7 turn that into attention. The specific combination isn't in the cited literature, as far as I can tell. The paper is clearly organized, and the tensor machinery is internally consistent.

The soft spots are real, and the stress-test concern holds up. p_ijt = o_it · o_jt is a bilinear form over MLP outputs; without any constraint or validation, it need not correspond to neighborhood overlap. It's a learnable attention score. The paper itself lists ablation studies as future work (Sec. V), so the mechanism is never isolated. There are also no error bars or significance tests, the text says three datasets while Table I lists two, and the HGNN backbone is cited only to the author's own prior paper [52]. These are not individually fatal, but together they mean Table II is not evidence for the headline claim.

That said, the results are not absurd—M7's improvements are consistent across both datasets and both metrics. If the authors added a generic attention baseline as an ablation, correlated p_ijt with actual common-neighbor counts, and released code, the claim could be tested. Without that, the novelty is mostly "attention-weighted HGNN," which is incremental.

Who this is for: practitioners who want a cheap high-order attention variant for dynamic link prediction, and referees who want to see the overlap claim confirmed or refuted.

Recommendation: send it to peer review—there is a real, fixable scientific question here—but the review should demand ablations and validation of the overlap score. If the authors can't show the score tracks overlap, the paper reduces to an incremental architecture tweak.

Referee Report

4 major / 4 minor

Summary. The paper proposes NO-HGNN for dynamic graph link prediction. It computes structural features o_it by passing B = sum_{k=1}^K A^k through a learnable MLP (Eqs. 3-5), defines a pairwise overlap score p_ijt = o_it · o_jt (Eq. 6), softmax-normalizes it (Eq. 7), and injects the resulting weights into a tensor-product HGNN (Eq. 2). The paper reports F1 and Accuracy on two datasets (ask-ubuntu and bitcoin-alpha, Table II) and claims improvements over six static and dynamic baselines. The central claim is that neighborhood-overlap modeling, not generic attention, drives the gains.

Significance. If the overlap mechanism were validated, the paper would offer a relevant and plausible extension of high-order message passing for dynamic graphs; the use of unnormalized adjacency powers (Eq. 3) and tensor formulations is a sensible design direction. Credit is due for targeting a real limitation of pairwise message-passing DGNNs. However, the manuscript does not establish that the learned score in Eq. (6) corresponds to neighborhood overlap, does not isolate the module by ablation, and provides no statistical support for the reported margins. These gaps are load-bearing for the stated contribution, so the result as presented is promising but unverified.

major comments (4)

[Section III-B, Eq. (6)] The central claim that p_ijt measures neighborhood overlap is not established. Since o_it is the output of a learnable MLP over B, the score o_it · o_jt is an unconstrained bilinear function of learned representations; it mixes contributions from common and non-common neighbors through cross terms and can be trained to encode arbitrary pairwise similarities that help link prediction. The paper offers no correlation check against a true overlap statistic (e.g., common-neighbor count or Jaccard coefficient), and Eq. (7) merely softmax-normalizes this learned score. As written, NO-HGNN is structurally a softmax-attention GNN over structural features, and the stated mechanism for the Table II gains is unsupported. Please add an ablation or a constrained formulation showing that overlap is actually captured.
[Section V(b)] The paper explicitly lists 'conducting ablation studies to isolate the contributions of neighborhood overlap estimation and high-order message passing' as future work. Because the overlap module is never removed, replaced, or compared against a backbone-only variant, the experiments cannot attribute the improvements over M6 to the overlap mechanism. At minimum, the revision should report variants with uniform/hop-based weights, true overlap scores in place of learned p_ijt, and no structural feature module.
[Section IV-A, Table I, and Table II] The text states that the study 'utilizes three DG datasets,' but Table I lists only D1 and D2, and Table II reports results only for those two datasets. This inconsistency must be corrected. With only two datasets and no diversity of scales or domains, the empirical support for general claims about dynamic graph learning is narrow.
[Table II and Section IV-B] No error bars, number of runs, or significance tests are reported. The observed gaps are modest in absolute terms (e.g., on D1, M7 exceeds M6 by 0.0054 in F1 and 0.0123 in Accuracy; on D2 by 0.0132 in F1 and 0.0159 in Accuracy), and these differences could easily fall within run-to-run variance for deep models. The 'notable improvements' claim requires repeated trials and a paired significance test, or at least standard deviations.

minor comments (4)

[Section II-A and Section III] Several typos should be fixed, including 'defination' (Section II-A), 'propsed' (Section V), and 'neighborhood overlay' (Section I and Section III).
[Equation (1)] Equation (1) does not render correctly: the tensor-product symbol and mode-3 product contain placeholder characters, and the role of the matrix M in the product is not explained. Readers cannot verify the tensor product definition without consulting reference [52].
[Section IV-A] The hyperparameter description gives ranges for the learning rate and L2 coefficient but does not report the negative sampling ratio, the early-stopping criterion in epochs, or the validation selection rule; these details are needed for reproducibility.
[Section IV-A, Baselines] For the static baselines M1 and M2, the manuscript does not explain how they are adapted to the dynamic setting (e.g., per-snapshot training or a pooled adjacency matrix), which is necessary to interpret their lower scores.

Circularity Check

1 steps flagged · score 4.0 of 10

Eq. (6) defines the 'overlap score' as a learned dot product, so the paper's claim that this score measures neighborhood overlap is definitional; the link-prediction gains themselves remain empirically grounded.

self definitional [Section III-B, Eq. (6)]
"To evaluate the impact of neighborhood overlap during message propagation, the overlap score between nodes i and j is computed as: p̂_ijt = o_it · o_jt (6) where p̂_ijt represents the degree of overlap between the neighborhoods of nodes."

No independent definition or ground-truth measure of 'neighborhood overlap' is given anywhere in the paper. The only formal content of the 'overlap score' is the dot product p̂_ijt = o_it · o_jt, where o_it is the output of a learned MLP over B (Eq. 5). Thus the assertion that p̂_ijt 'represents the degree of overlap' and the later claim that M7's accuracy gains 'stem from M7's ability to adopt neighbor overlap correlations' reduce, by construction, to saying that the model uses a learned dot-product similarity — a generic attention mechanism. The overlap label adds no independently verified structural content, and Section V itself defers 'conducting ablation studies to isolate the contributions of neighborhood overlap estimation' to future work.

full rationale

The empirical evaluation is self-contained: NO-HGNN is trained on labeled training pairs and evaluated on a held-out test split against six baselines, so the accuracy and F1 improvements in Table II are not fitted inputs renamed as predictions. The circularity risk sits in the mechanism label. Equation (6) defines the overlap score as a dot product of learned structural features, and no external or internal measure of neighborhood overlap is introduced to validate that interpretation. Consequently, the claim that the model is 'neighborhood overlap-aware' is true only by definition: the overlap score is whatever the learned dot product is, making the proposed module a softmax-attention mechanism over learned structural features. The HGNN backbone is cited to the author's own prior work [52], but Eq. (2) is stated in full and is a standard message-passing update, so that self-citation is not load-bearing for the empirical results. Overall, one definitional/renaming step sits at the core of the claimed novelty, giving partial circularity, while the benchmark comparison retains independent content.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The model is a standard supervised GNN; almost all predictive capacity comes from trained parameters and hand-chosen hyperparameters. The only conceptual novelty is the claimed equivalence between a dot product of learned features and neighborhood overlap, which is asserted without evidence.

free parameters (6)

K (max hop count) = not reported
Eq. (3) sums adjacency powers up to hop K; K is never specified in the experiments and is not ablated.
feature dimension F = 32
Set to 32 for all models; no sensitivity analysis is given.
number of HGNN layers l = 2
Fixed to 2; no justification is provided.
L2 regularization coefficient beta = chosen from {0.01, 0.005, 0.001, 0.0005}; final value not reported
Grid-searched per model; the selected value does not appear in the results.
learning rate = chosen from {0.1, 0.01, 0.02, 0.05, 0.001, 0.002}; final value not reported
Grid-searched; final value not reported.
negative sampling ratio = not reported
Training data construction uses negative sampling following [58], but the ratio is not given.

assumptions (6)

standard math Tensor face-wise product and mode-3 product are well-defined and computed as in [52].
Used in Eqs. (1)-(2) for the HGNN backbone; no proof is provided but the operations are standard.
domain assumption All graph snapshots share the same node set V with constant size N.
Problem definition in Section II-A assumes a fixed node set across time; real dynamic graphs often add or remove nodes, and no preprocessing handles this.
ad hoc to paper The dot product of structural features o_it dot o_jt represents neighborhood overlap.
Introduced in Eq. (6) without validation or comparison to established overlap measures.
ad hoc to paper Summing adjacency powers up to K holds multi-hop neighborhood overlap information.
Eq. (3) assumes raw k-hop path counts are informative for overlap; K is not reported.
domain assumption Negative sampling yields a valid training distribution for link prediction.
Section IV-A applies negative sampling following [58], but the ratio and procedure are unspecified.
ad hoc to paper Softmax normalization of overlap scores produces useful aggregation weights.
Eq. (7) uses softmax; this is a typical attention choice, but its contribution is not ablated.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Neighborhood Overlap-Aware High-Order Graph Neural Network for Dynamic Graph Learning." pith.science (2026). https://pith.science/paper/A5MVBBD7

@misc{pith2026250606728,
  author       = {Pith},
  title        = {Pith review of: Neighborhood Overlap-Aware High-Order Graph Neural Network for Dynamic Graph Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/A5MVBBD7}},
  note         = {Machine review of arXiv:2506.06728}
}

read the original abstract

Dynamic graph learning (DGL) aims to learn informative and temporally-evolving node embeddings to support downstream tasks such as link prediction. A fundamental challenge in DGL lies in effectively modeling both the temporal dynamics and structural dependencies of evolving graph topologies. Recent advances in Dynamic Graph Neural Networks (DGNNs) have obtained remarkable success by leveraging message-passing mechanisms to capture pairwise node interactions. However, these approaches often overlook more complex structural patterns, particularly neighborhood overlap, which can play a critical role in characterizing node interactions. To overcome this limitation, we introduce the Neighborhood Overlap-Aware High-Order Graph Neural Network (NO-HGNN), which is built upon two key innovations: (a) computing a correlation score based on the extent of neighborhood overlap to better capture complex node interactions; and (b) embedding this correlation directly into the message-passing process of high-order graph neural networks in the DGL. Experiments on two real-world dynamic graphs show that NO-HGNN achieves notable improvements in link prediction accuracy, outperforming several state-of-the-art approaches.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

68 extracted references · 62 canonical work pages

[52]

GT-A2T: Graph tensor alliance attention network ,

L. Wang, K . Liu, and Y. Yuan, “GT-A2T: Graph tensor alliance attention network ,” IEEE/CAA Journal of Automatica Sinica , DOI: 10.1109/JAS.2024.124863, 2024

work page arXiv 2024
[1]

Performance of latent factor models with extended linear biases,

J. Chen, X. Luo, Y. Yuan, M. Shang, Z. Ming, Z. Xiong, “Performance of latent factor models with extended linear biases,” Knowledge-Based Systems, vol. 123, pp. 128-136, 2017

work page 2017
[2]

Semi-supervised classification with graph convolutional networks,

T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” in Proc. of the Int. Conf. on Learning Representations, pp. 1-14, 2017

work page 2017
[3]

Learning conjoint attentions for graph neural nets,

T. He, Y. S. Ong, and L. Bai, “Learning conjoint attentions for graph neural nets,” in Proc. of Conf. on Neural Information Processing Systems, 2021

work page 2021
[4]

Enhancing graph convolutional networks with an efficient k-hop neighborhood approach,

J. Chen, X. Luo, Y. Yuan, Z. Wang, “Enhancing graph convolutional networks with an efficient k-hop neighborhood approach,” Information Fusion, vol. 124, 2025

work page 2025
[5]

Attention-mechanism-based neural latent-factorization-of-tensors model,

X. Xu, M. Lin, Z. Xu, and X. Luo, “Attention-mechanism-based neural latent-factorization-of-tensors model,” ACM Trans. Knowl. Discov. Data, pp. 1- 27, 2025

work page 2025
[6]

A generalized and fast -converging non -negative latent factor model for predicting user preference s in recommender systems,

Y. Yuan, X. Luo, M. Shang, and D. Wu, “A generalized and fast -converging non -negative latent factor model for predicting user preference s in recommender systems,” in Proc.of The ACM Web Conf. pp. 498-507, 2020

work page 2020
[7]

Graph linear convolution pooling for learning in incomplete high -dimensional data,

F. Bi, T. He, Y. S. Ong, and X. Luo, “Graph linear convolution pooling for learning in incomplete high -dimensional data,” IEEE Trans. on Knowledge and Data Engineering, vol. 37, no. 4, pp. 1838-1852, 2025

work page 2025

Show all 68 references

[8]

A stochastic approach to analysis of energy-aware DVS-enabled cloud datacenters,

Y. Xia, M. Zhou, X. Luo, S. Pang and Q. Zhu, “A stochastic approach to analysis of energy-aware DVS-enabled cloud datacenters,” IEEE Transactions on Systems, Man, and Cybernetics: Systems, vol. 45, no. 1, pp. 73-83, 2015

2015
[9]

Hyper-parameter-evolutionary latent factor analysis for high -dimensional and sparse data from recommender systems,

J. Chen, Y . Yuan, T . Ruan, J . Chen, and X . Luo, “Hyper-parameter-evolutionary latent factor analysis for high -dimensional and sparse data from recommender systems,” Neurocomputing, 2020, 421: 316-328

2020
[10]

Spatial-temporal tensor graph convolutional network for traffic speed prediction,

X. Xu, T. Zhang, C. Xu, Z. Cui, and J. Yang, “Spatial-temporal tensor graph convolutional network for traffic speed prediction,” IEEE Trans. on Intelligent Transportation Systems, vol. 24, no. 1, pp. 92-103, 2023

2023
[11]

An L1-and-L2-norm-oriented latent factor model for recommender systems,

D. Wu, M. Shang, X. Luo, and Z. Wang, “An L1-and-L2-norm-oriented latent factor model for recommender systems,” IEEE Trans. on Neural Networks and Learning Systems, vol. 33, no. 10, pp. 5775-5788, 2022

2022
[12]

Adaptive divergence-based non-negative latent factor analysis of high -dimensional and incomplete matrices from industrial applications,

Y. Yuan, X. Luo, and M . Zhou, “Adaptive divergence-based non-negative latent factor analysis of high -dimensional and incomplete matrices from industrial applications,” IEEE Trans. on Emerging Topics in Computational Intelligence, vol. 8, no.2, pp. 1209-1222. 2024

2024
[13]

A prediction -sampling-based multilayer -structured latent factor model for accurate representation of high - dimensional and sparse data,

D. Wu, X. Luo, Y. He, and M. Zhou, “A prediction -sampling-based multilayer -structured latent factor model for accurate representation of high - dimensional and sparse data,” IEEE Trans. on Neural Networks and Learning Systems, vol. 35, no. 3, pp. 3845-3858, 2024

2024
[14]

A Novel approach to large-scale dynamically weighted directed network representation,

X. Luo, H. Wu, Z. Wang, J. Wang, and D. Meng, “A Novel approach to large-scale dynamically weighted directed network representation,” IEEE Trans. on Pattern Analysis and Machine Intelligence, vol. 44, no. 12, pp. 9756-9773, 2022

2022
[15]

Effects of preprocessing and training biases in latent factor models for recommender systems,

Y. Yuan, X. Luo, and M. Shang, “Effects of preprocessing and training biases in latent factor models for recommender systems,” Neurocomputing, pp. 2019-2030, 2018

2019
[16]

Two-stream graph convolutional network-incorporated latent feature analysis,

F. Bi, T. He, Y. Xie, and X. Luo, “Two-stream graph convolutional network-incorporated latent feature analysis,” IEEE Trans. on Services Computing, vol. 16, no. 4, pp. 3027-3042, 2023

2023
[17]

A novel machine learning system for industrial robot arm calibration,

Z. Li, S. Li and X. Luo, “A novel machine learning system for industrial robot arm calibration,” IEEE Trans. on Circuits and Systems II: Express Briefs, vol. 71, no. 4, pp. 2364-2368, 2024

2024
[18]

Neulft: A novel approach to nonlinear canonical polyadic decomposition on high -dimensional incomplete tensors,

X. Luo, H. Wu and Z. Li, “Neulft: A novel approach to nonlinear canonical polyadic decomposition on high -dimensional incomplete tensors,” IEEE Trans. on Knowledge and Data Engineering, vol. 35, no. 6, pp. 6148-6166, 2023

2023
[19]

Proximal alternating-direction-method-of-multipliers-incorporated nonnegative latent factor analysis,

F. Bi, X. Luo, B. Shen, H. Dong, and Z. Wang, “Proximal alternating-direction-method-of-multipliers-incorporated nonnegative latent factor analysis,” IEEE/CAA Journal of Automatica Sinica, vol. 10, no. 6, pp. 1388-1406, 2023

2023
[20]

A proportional integral controller-enhanced non-negative latent factor analysis model, IEEE/CAA Journal of Automatica Sinica,

Y. Yuan, S. Lu, and X. Luo, “A proportional integral controller-enhanced non-negative latent factor analysis model, IEEE/CAA Journal of Automatica Sinica,” DOI: 10.1109/JAS.2024.125055, 2024

2024
[21]

Temporal pattern-aware QoS prediction via biased non-negative latent factorization of tensors,

X. Luo, H. Wu, H. Yuan, and M. Zhou, “Temporal pattern-aware QoS prediction via biased non-negative latent factorization of tensors,” IEEE Trans. on Cybernetics, vol. 50, no. 5, pp. 1798-1809, 2020

2020
[22]

A nonlinear PID-incorporated adaptive stochastic gradient descent algorithm for latent factor analysis,

J. Li, X. Luo, Y. Yuan, and S. Gao, “A nonlinear PID-incorporated adaptive stochastic gradient descent algorithm for latent factor analysis,” IEEE Trans. on Automation Science and Engineering, vol. 21, no. 3, pp. 3742-3756, 2024

2024
[23]

Polarized message-passing in graph neural networks,

T. He, Y. Liu, Y. S. Ong, X. Wu, and X. Luo, “Polarized message-passing in graph neural networks,” Artificial Intelligence, vol. 331, p. 104129, 2024

2024
[24]

Po sition-transitional particle swarm optimization-incorporated latent factor analysis,

X. Luo, Y. Yuan, S. Chen, N. Zeng, and Z. Wang, “Po sition-transitional particle swarm optimization-incorporated latent factor analysis,” IEEE Trans. on Knowledge and Data Engineering, vol. 34, no. 8, pp. 3958-3970, 2022

2022
[25]

Latent-Factorization-of-Tensors-Incorporated Battery Cycle Life Prediction,

M. Chen, L. Tao, J. Lou, and X. Luo, “Latent-Factorization-of-Tensors-Incorporated Battery Cycle Life Prediction,” IEEE/CAA Journal of Automatica Sinica, vol. 12, no. 3, pp. 633-635, 2025

2025
[26]

An adaptive divergence -based non -negative latent factor model,

Y. Yuan, R. Wang, G. Yuan, and X. Luo, “An adaptive divergence -based non -negative latent factor model,” IEEE Trans. on Systems, Man, and Cybernetics: Systems, vol. 53, no. 10, pp. 6475-6487, 2023

2023
[27]

LightGCN: Simplifying and powering rgraph convolution network for recommendation

X. He, K. Deng, X. Wang, Y. Li, Y. Zhang, and M. Wang, “LightGCN: Simplifying and powering rgraph convolution network for recommendation”, in Proc. of the ACM SIGIR Conf. on Research and Development in Information Retrieval, pp. 639-648, 2020

2020
[28]

A proportional -integral-derivative-incorporated stochastic gradient descent-based latent factor analysis model,

J. Li, Y. Yuan, T. Ruan, J. Chen, and X. Luo, “A proportional -integral-derivative-incorporated stochastic gradient descent-based latent factor analysis model,” Neurocomputing, vol. 427, pp. 29-39, 2021

2021
[29]

A Latent Factor Analysis-Based Approach to Online Sparse Streaming Feature Selection,

D. Wu, Y. He, X. Luo, and M. Zhou, "A Latent Factor Analysis-Based Approach to Online Sparse Streaming Feature Selection," IEEE Transactions on Systems, Man, and Cybernetics: Systems, vol. 52, no. 11, pp. 6744-6758, 2022

2022
[30]

A highly-accurate three-way decision-incorporated online sparse streaming features selection model,

R. Xu, D. Wu, R. Wang and X. Luo, “A highly-accurate three-way decision-incorporated online sparse streaming features selection model,” IEEE Trans. on Systems, Man, and Cybernetics: Systems, doi: 10.1109/TSMC.2025.3548648, 2025

2025
[31]

A multilayered-and-randomized latent factor model for high-dimensional and sparse matrices,

Y. Yuan, Q. He, X. Luo, and M. Shang, “A multilayered-and-randomized latent factor model for high-dimensional and sparse matrices,” IEEE Trans. on Big Data, vol. 8, no. 3, pp. 784-794, 2022

2022
[32]

Robust Low-Rank Latent Feature Analysis for Spatiotemporal Signal Recovery,

D. Wu, Z. Li, Z. Yu, Y. He, and X. Luo, "Robust Low-Rank Latent Feature Analysis for Spatiotemporal Signal Recovery," IEEE Transactions on Neural Networks and Learning Systems, vol. 36, no. 2, pp. 2829-2842, 2025

2025
[33]

An α-β-divergence-generalized recommender for highly accurate predictions of missing user preferences,

M. Shang, Y. Yuan, X. Luo, and M. Zhou, “An α-β-divergence-generalized recommender for highly accurate predictions of missing user preferences,” IEEE Trans. on Cybernetics, vol. 52, no. 8, pp. 8006-8018, 2022

2022
[34]

A deep latent factor model for high-dimensional and sparse matrices in recommender systems,

D. Wu, X. Luo, M. Shang, Y. He, G. Wang and M. Zhou, “A deep latent factor model for high-dimensional and sparse matrices in recommender systems,” IEEE Trans. on Systems, Man, and Cybernetics: Systems, vol. 51, no. 7, pp. 4285-4296, 2021

2021
[35]

Local Search -Based Anytime Algorithms for Continuous Distributed Constraint Optimization Problems,

X. Liao, K. Hoang and X. Luo, “Local Search -Based Anytime Algorithms for Continuous Distributed Constraint Optimization Problems,” IEEE/CAA Journal of Automatica Sinica, vol. 12, no. 1, pp. 288-290, 2025

2025
[36]

A highly accurate framework for self -labeled semisupervised classification in industrial applications,

D. Wu, X. Luo, G. Wang, M. Shang, Y. Yuan and H. Yan, “A highly accurate framework for self -labeled semisupervised classification in industrial applications,” IEEE Trans. on Industrial Informatics, vol. 14, no. 3, pp. 909-920, 2018

2018
[37]

Learning error re finement in stochastic gradient descent-based latent factor analysis via diversified pid controllers,

J. Li, Y. Yuan, and X. Luo, “Learning error re finement in stochastic gradient descent-based latent factor analysis via diversified pid controllers,” IEEE Trans. on Emerging Topics in Computational Intelligence, doi: 10.1109/TETCI.2025.3547854
[38]

MKG -FENN: a multimodal knowledge graph fused end -to-end neural network for accurate drug -drug interaction prediction,

D. Wu, W. Sun, Y. He, Z. Chen, and X. Luo, “MKG -FENN: a multimodal knowledge graph fused end -to-end neural network for accurate drug -drug interaction prediction,” in Proc. of the AAAI Conf. on Artificial Intelligence, pp. 10216-10224, 2024

2024
[39]

Dynamic graph convolutional networks,

F. Manessi, A. Rozza, and M. Manzo, “Dynamic graph convolutional networks,” Pattern Recognition, vol. 97, 2020

2020
[40]

EvolveGCN: Evolving graph convolutional networks for dynamic graphs,

A. Pareja, G. Domeniconi, J. Chen, T. Ma, T. Suzumura, H. Kanezashi, T. Kaler, T. B. Schardl, C. E. Leiserson, “EvolveGCN: Evolving graph convolutional networks for dynamic graphs,” in Proc. of the AAAI Conf. on Artificial Intelligence, pp. 5363-5370, 2020

2020
[41]

Temporal web service QoS prediction via Kalman filter -incorporated dynamic latent factor analysis,

Y. Yuan, M. Shang, and X . Luo, “Temporal web service QoS prediction via Kalman filter -incorporated dynamic latent factor analysis,” in Proc of European Conf. on Artificial Intelligence, pp. 561-568, 2020

2020
[42]

An Adaptive p-Norms-Based Kinematic Calibration Model for Industrial Robot Positioning Accuracy Promotion,

T. Chen, W. Yang, S. Li, and X. Luo, “An Adaptive p-Norms-Based Kinematic Calibration Model for Industrial Robot Positioning Accuracy Promotion,” IEEE Transactions on Systems, Man, and Cybernetics: Systems, vol. 55, no. 4, pp. 2937-2949, 2025

2025
[43]

Online sparse streaming feature selection via decision risk,

R. Xu, D. Wu, and X. Luo, “Online sparse streaming feature selection via decision risk,” 2023 IEEE International Conference o n Systems, Man, and Cybernetics, pp. 4190-4195, 2023

2023
[44]

A State-migration particle swarm optimizer for adaptive latent factor analysis of high-dimensional and incomplete data,

J. Chen, K. Liu, X. Luo, Y. Yuan, K. Sedraoui, Y. Al-Turki, and M. Zhou, “A State-migration particle swarm optimizer for adaptive latent factor analysis of high-dimensional and incomplete data,” IEEE/CAA Journal of Automatica Sinica, vol. 11, no. 11, pp. 2220-2235, 2024

2024
[45]

A fast nonnegative autoencoder -based approach to latent feature analysis on high -dimensional and incomplete data,

F. Bi, T. He, and X. Luo, “A fast nonnegative autoencoder -based approach to latent feature analysis on high -dimensional and incomplete data,” IEEE Trans. on Services Computing, vol. 17, no. 3, pp. 733-746, 2024

2024
[46]

A Kalman-filter-incorporated latent factor analysis model for temporally dynamic sparse data,

Y. Yuan, X. Luo, M. Shang, and Z. Wang, “A Kalman-filter-incorporated latent factor analysis model for temporally dynamic sparse data,” IEEE Trans. on Cybernetics, vol. 53, no. 9, pp. 5788-5801, 2023

2023
[47]

DySAT: deep neural representation learning on dynami c graphs via self -attention networks,

A. Sankar, Y. Wu, L. Gou, W. Zhang, and H. Yang, “DySAT: deep neural representation learning on dynami c graphs via self -attention networks,” in Proc. of the Int. Conf. on Web Search and Data Mining, pp. 519-527, 2020

2020
[48]

Pseudo gradient-adjusted particle swarm optimization for accurate adaptive latent factor analysis,

X. Luo, J. Chen, Y. Yuan and Z. Wang, “Pseudo gradient-adjusted particle swarm optimization for accurate adaptive latent factor analysis,” IEEE Trans. on Systems, Man, and Cybernetics: Systems, vol. 54, no. 4, pp. 2213-2226, 2024

2024
[49]

Neural tucker factorization,

P. Tang and X. Luo, “Neural tucker factorization,” IEEE/CAA Journal of Automatica Sinica, vol. 12, no. 2, pp. 475-477, 2025

2025
[50]

Non-negative latent factor model based on β-divergence for recommender systems,

X. Luo, Y. Yuan, M. Zhou, Z. Liu , and M. Shang, “Non-negative latent factor model based on β-divergence for recommender systems,” IEEE Trans. Systems, Man, and Cybernetics: Systems, vol. 51, no. 8, pp. 4612-4623, 2021

2021
[51]

An outlier-resilient autoencoder for representing high-dimensional and incomplete data,

D. Wu, Y. Hu, K. Liu, J. Li, X. Wang, S. Deng, N. Zheng, and X. Luo, “An outlier-resilient autoencoder for representing high-dimensional and incomplete data,” IEEE Trans. on Emerging Topics in Computational Intelligence, doi: 10.1109/TETCI.2024.3437370, 2024
[53]

Symmetry -constrained Non-negative Matrix Factorization Approach for Highly -Accurate Community Detection,

Z. Liu, X. Luo and M. Zhou, "Symmetry -constrained Non-negative Matrix Factorization Approach for Highly -Accurate Community Detection," 2021 IEEE 17th International Conference on Automation Science and Engineering, 2021, pp. 1521-1526

2021
[54]

Self-training semi -supervised classification based on density peaks of data,

D. Wu, M. Shang, X. Luo, J. Xu, H. Yan, W. Deng, G. Wang, “Self-training semi -supervised classification based on density peaks of data,” Neurocomputing, Vol. 275, pp.180-191, 2018

2018
[55]

SDGNN: symmetry-preserving dual-stream graph neural networks,

J. Chen, Y. Yuan, and X. Luo, “SDGNN: symmetry-preserving dual-stream graph neural networks,” IEEE/CAA Journal of Automatica Sinica, no. 11, vol. 7, 1717-1719, 2024

2024
[56]

Accelerated Non-negative Latent Factor Analysis on High -Dimensional and Sparse Matrices via Generalized Momentum Method,

Z. Liu, X. Luo, S. Li and M. Shang, "Accelerated Non-negative Latent Factor Analysis on High -Dimensional and Sparse Matrices via Generalized Momentum Method," 2018 IEEE International Conference on Systems, Man, and Cybernetics, 2018, 3051-3056

2018
[57]

Improving network topology -based protein interactome mapping via collaborative filtering,

X. Luo, Z . Ming, Z . You, S . Li, Y . Xia, H . Leung, “Improving network topology -based protein interactome mapping via collaborative filtering, ” Knowledge-Based Systems, vol. 90, pp. 23-32, 2015

2015
[58]

Recursion -and-Fuzziness Reinforced Online Sparse Streaming Feature Selection,

R. Xu, D. Wu and X. Luo, “Recursion -and-Fuzziness Reinforced Online Sparse Streaming Feature Selection,” IEEE Trans. on Fuzzy Systems, doi: 10.1109/TFUZZ.2025.3569272

2025
[59]

Algorithms of unconstrained non-negative latent factor analysis for recommender systems,

X. Luo, M. Zhou, S. Li, D. Wu, Z. Liu and M. Shang, “Algorithms of unconstrained non-negative latent factor analysis for recommender systems,” IEEE Transactions on Big Data, vol. 7, no. 1, pp. 227-240, 2021

2021
[60]

Adaptively-accelerated parallel stochastic gradient descent for high -dimensional and incomplete data representation learning,

W. Qin, X. Luo and M. Zhou, “Adaptively-accelerated parallel stochastic gradient descent for high -dimensional and incomplete data representation learning,” IEEE Transactions on Big Data, vol. 10, no. 1, pp. 92-107, 2024

2024
[61]

Symmetry and nonnegativity-constrained matrix factorization for community detection,

Z. Liu, G. Yuan, and X. Luo, “Symmetry and nonnegativity-constrained matrix factorization for community detection,” IEEE/CAA Journal of Automatica Sinica, vol. 9, no. 9, pp. 1691-1693, 2022

2022
[62]

Randomized latent factor model for high -dimensional and sparse matrices from industrial applications,

M. Shang, X. Luo, Z. Liu, J. Chen, Y. Yuan and M. Zhou, “Randomized latent factor model for high -dimensional and sparse matrices from industrial applications,” IEEE/CAA Journal of Automatica Sinica, vol. 6, no. 1, pp. 131-141, 2019

2019
[63]

A node-collaboration-informed graph convolutional network for highly accurate representation to undirected weighted graph,

Y. Yuan, Y. Wang, and X. Luo, “A node-collaboration-informed graph convolutional network for highly accurate representation to undirected weighted graph,” IEEE Trans. on Neural Networks and Learning Systems, DOI: 10.1109/TNNLS.2024.3514652, 2024
[64]

Symmetry and graph bi-regularized non-negative matrix factorization for precise community detection,

Z. Liu, X. Luo and M. Zhou, “Symmetry and graph bi-regularized non-negative matrix factorization for precise community detection,” IEEE Transactions on Automation Science and Engineering, vol. 21, no. 2, pp. 1406-1420, 2024

2024
[65]

Vibration Control using A Robust Input Shaper via Extended Kalman Filter-Incorporated Residual Neural Network,

W. Yang, S. Li and X. Luo, “Vibration Control using A Robust Input Shaper via Extended Kalman Filter-Incorporated Residual Neural Network,” IEEE International Conference on Systems, Man, and Cybernetics, 2024, pp. 3980-3986

2024
[66]

A two-stream light graph convolution network -based latent factor model for accurate cloud service qos estimation ,

F. Bi, T. He and X. Luo, “A two-stream light graph convolution network -based latent factor model for accurate cloud service qos estimation ,” IEEE International Conference on Data Mining, 2022, pp. 855-860

2022
[67]

Analysis of students’ positive emotion and smile intensity using sequence-relative key-frame labeling and deep-asymmetric convolutional neural network,

Z. Luo, X. Jin, Y. Luo, Q. Zhou, and X. Luo, “Analysis of students’ positive emotion and smile intensity using sequence-relative key-frame labeling and deep-asymmetric convolutional neural network,” IEEE/CAA J. Autom. Sinica, vol. 12, no. 4, pp. 806–820, 2025

2025
[68]

Constraint-induced symmetric nonnegative matrix factorization for accurate community detection,

Z. Liu, X. Luo, Z. Wang, X. Liu, “Constraint-induced symmetric nonnegative matrix factorization for accurate community detection,” Information Fusion, vol. 89, pp. 588-602, 2023

2023

Pith tools

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

[1] [52]

GT-A2T: Graph tensor alliance attention network ,

L. Wang, K . Liu, and Y. Yuan, “GT-A2T: Graph tensor alliance attention network ,” IEEE/CAA Journal of Automatica Sinica , DOI: 10.1109/JAS.2024.124863, 2024

work page arXiv 2024

[2] [1]

Performance of latent factor models with extended linear biases,

J. Chen, X. Luo, Y. Yuan, M. Shang, Z. Ming, Z. Xiong, “Performance of latent factor models with extended linear biases,” Knowledge-Based Systems, vol. 123, pp. 128-136, 2017

work page 2017

[3] [2]

Semi-supervised classification with graph convolutional networks,

T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” in Proc. of the Int. Conf. on Learning Representations, pp. 1-14, 2017

work page 2017

[4] [3]

Learning conjoint attentions for graph neural nets,

T. He, Y. S. Ong, and L. Bai, “Learning conjoint attentions for graph neural nets,” in Proc. of Conf. on Neural Information Processing Systems, 2021

work page 2021

[5] [4]

Enhancing graph convolutional networks with an efficient k-hop neighborhood approach,

J. Chen, X. Luo, Y. Yuan, Z. Wang, “Enhancing graph convolutional networks with an efficient k-hop neighborhood approach,” Information Fusion, vol. 124, 2025

work page 2025

[6] [5]

Attention-mechanism-based neural latent-factorization-of-tensors model,

X. Xu, M. Lin, Z. Xu, and X. Luo, “Attention-mechanism-based neural latent-factorization-of-tensors model,” ACM Trans. Knowl. Discov. Data, pp. 1- 27, 2025

work page 2025

[7] [6]

A generalized and fast -converging non -negative latent factor model for predicting user preference s in recommender systems,

Y. Yuan, X. Luo, M. Shang, and D. Wu, “A generalized and fast -converging non -negative latent factor model for predicting user preference s in recommender systems,” in Proc.of The ACM Web Conf. pp. 498-507, 2020

work page 2020

[8] [7]

Graph linear convolution pooling for learning in incomplete high -dimensional data,

F. Bi, T. He, Y. S. Ong, and X. Luo, “Graph linear convolution pooling for learning in incomplete high -dimensional data,” IEEE Trans. on Knowledge and Data Engineering, vol. 37, no. 4, pp. 1838-1852, 2025

work page 2025