Pith. sign in

REVIEW 5 major objections 7 minor 36 references

Consistency-Aware Padding for Incomplete Multi-Modal Alignment Clustering Based on Self-Repellent Greedy Anchor Search

T0 review · 5 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read CAPIMAC claims that incomplete and misaligned multimodal data can be clustered by padding the shorter view with equal-weight Gaussian-kernel interpolations of the worst-matched aligned pairs and re-running Hungarian alignment.

desk verdict Plausible incremental method for incomplete-and-misaligned multi-view clustering; consistent gains on small benchmarks, but the core padding assumption is untested and the empirical reporting is thin. read the letter →

arxiv 2507.03917 v1 pith:Q65J7UDH submitted 2025-07-05 cs.LG cs.CV

Shubin Ma , Liang Zhao , Mingdong Lu , Yifan Guo , Bo Xu This is my paper · ORCID
classification cs.LGcs.CV
keywords incompletemultimodalclusteringpartialviewalignmentanchor-basedself-repellentrandomwalkGaussiankernelinterpolationnoise-contrastivelearningHungariandatafusion
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 is trying to establish that the hard case of multimodal clustering -- data that are simultaneously incomplete (some samples missing) and misaligned (view order shuffled) -- can be handled by creating virtual samples rather than by discarding or class-level aligning the available data. Its method, CAPIMAC, first finds anchors with a self-repellent greedy search, learns noise-robust latent features, and then pads the shorter view using equal-weight Gaussian-kernel interpolation between the most dissimilar already-aligned pairs before applying Hungarian alignment. If the claim is right, fusion quality no longer depends on having the same number of matched samples in each view, and incomplete misaligned data can be clustered directly. The supporting evidence is that CAPIMAC reports the best accuracy, NMI, ARI, and F1 in the majority of comparisons across five benchmark datasets at alignment rates 0.3, 0.5, and 0.7 with 50 percent missing data.

What carries the argument

The load-bearing machinery is the consistency-aware padding module (CAPM) built on anchor-based re-representation and noise-contrastive training. Anchors are chosen by a self-repellent random walk transition kernel $P^{(v)}_{ij}$ with decay $r_{\mu_i}(x_i)=(x_i/\mu_i)^{-\alpha}$ ($\alpha=0.5$) plus a greedy radius-$d$ rule that guarantees every class contributes an anchor, and the original data are re-represented as $\bar{X}=X A^\top$. Training uses a cosine-distance noise-contrastive loss that down-weights false negative pairs. For padding, CAPM computes the Hungarian distance matrix $Z=1-Z'$, reorders rows to $Z^r=U^r Z$, finds the $n_k=n_l-n_s$ pairs with the largest similarity gap, and synthesizes missing samples by equal-weight Gaussian kernel interpolation $$y_{\text{target}}=\sum_{i=1}^{n_s}\frac{K(x_{\text{target}},x_i,\$\sigma$)}{\sum_{j=1}^{n_s}K(x_{\text{target}},x_j,\$\sigma$)}y_i,$$ which equalizes view sizes so Hungarian alignment can run on an $n_l \times n_l$ matrix.

What would settle it

Take a fixed incomplete-misaligned split and run CAPIMAC with padding replaced by interpolation between random pairs instead of the most dissimilar pairs; if clustering metrics stay essentially unchanged, the specific dissimilarity-based padding rule is not what drives the reported gains. More directly, measure the class purity of the virtual samples generated by Eq. (19) on a labeled benchmark, because a substantial fraction of padded points falling inside another class's cluster would falsify the consistency assumption.

Watch

Extended reading notes

Core claim

The paper's central claim is that the misalignment caused by missing data can be repaired by padding rather than by matching only. The consistency-aware padding module reorders the distance matrix to $Z^r = U^r Z$, takes the $n_k = n_l - n_s$ index pairs with the largest similarity differences, creates virtual samples by equal-weight Gaussian kernel interpolation, and then performs Hungarian alignment on the equalized $n_l \times n_l$ matrix. The claim is that this keeps data-pair similarity coherent and yields one-to-one alignments that preserve class structure, which is why the method reports the best result in the majority of comparisons over five datasets at alignment rates 0.3, 0.5, and 0.7 with 50 percent missing data. The paper further claims that self-repellent greedy anchors capture both local structure and global class coverage, and that a cosine-based noise-contrastive loss reduces false negatives.

Load-bearing premise

The padding module assumes that a virtual sample created by interpolating between the most dissimilar already-aligned samples of a view still lies inside the correct class, so that padding before re-alignment adds signal rather than injecting wrong-class noise.

Editorial extensions

If this is right

  • Incomplete views no longer have to be discarded: every retained sample can contribute to fusion, so data utilization rises when incompleteness is the problem.
  • Hungarian alignment operates on equal-size padded matrices, which removes the imbalance obstruction to one-to-one matching between views.
  • A cosine-based noise-contrastive loss reduces the influence of false-negative pairs, making the learned representation cleaner when the view order is shuffled.
  • Because anchors are selected from complete data, performance degrades at low alignment rates but still leads the comparison at rate 0.3 in most metrics, indicating the method tolerates alignment uncertainty.

Reading between the lines

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

  • Beyond the paper, the method's core padding assumption can be tested in isolation: measure how often the virtual samples from Eq. (19) receive the class label of their target pair, and if that purity is low, the reported gain would have to come from some other component.
  • Beyond the paper, because the anchor search relies on complete data, a natural extension is to select anchors over the union of available samples across views, which would push the method to very low alignment rates.
  • Beyond the paper, the padding-then-realign step is modular and could be grafted onto other Hungarian-based partial-alignment pipelines, a transfer the paper itself does not demonstrate.
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.

Referee Report

5 major / 7 minor

Summary. The paper proposes CAPIMAC, a method for clustering multimodal data that is simultaneously incomplete (missing samples in one or more modalities) and misaligned (no known one-to-one correspondence between modalities). The method combines a self-repellent random walk with a greedy anchor selection module to re-represent the data, a noise-contrastive loss for representation learning, and a consistency-aware padding module that uses Gaussian kernel interpolation to equalize modality sizes before Hungarian alignment. Experiments on five benchmark datasets compare CAPIMAC against seven baselines at alignment rates 0.3, 0.5 and 0.7 with a fixed incompleteness rate of 0.5, and an ablation study isolates the effect of the padding module.

Significance. If the proposed mechanism works as intended, the paper addresses a genuine and under-studied setting: multimodal clustering where incompleteness and misalignment co-occur. The experimental scope is reasonable, with five datasets, four metrics, seven baselines, and an ablation of the central module. Credit is due for the public code release promise, the breadth of the comparison, and for explicitly acknowledging a limitation of the anchor-selection step at low alignment rates. However, the central novelty is the padding heuristic, and its core assumption—that interpolating between the least similar aligned pairs preserves class structure—is neither justified nor directly validated. The reported results are single point estimates without variance or significance testing. The contribution is potentially valuable, but the evidence as presented is not yet sufficient to establish the claimed superiority.

major comments (5)
  1. [§3.3, Eq. (19)-(20)]
  2. [§3.1, Eq. (4)]
  3. [§4.2, Tables 1, 2, 4]
  4. [§4.2, fourth bullet]
  5. [§3.3, Eq. (16)]
minor comments (7)
  1. [§3.1, Eq. (7)]
  2. [Table 1]
  3. [Tables 1-4]
  4. [References]
  5. [§4.1]
  6. [§3.3, Eq. (19)]
  7. [§4.3]

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity found: the central claim is an empirical benchmark comparison against external baselines, and the paper's self-cited methods appear only as baselines, not as premises.

full rationale

The paper's central claim is that CAPIMAC achieves the best clustering performance in the majority of cases across five public datasets and three alignment rates. This claim is supported by tables comparing four metrics against seven published methods, including the authors' own EGPVC and DGPPVC. Those self-cited methods are baselines being outperformed, not premises from which the result is derived, so the self-citations are not load-bearing. The derivation chain for the proposed components (SRGASM, noise-contrastive loss, CAPM) does not contain any equation that, by construction, equals its own input. In particular, the CAPM interpolation in Eq. 19 does not predict a quantity that is fitted from the same data and then reported as a finding; instead, the ablation in Table 3 shows that adding the padding module improves clustering metrics relative to discarding missing data, and the final evaluation is against fixed external benchmarks. The phrase about virtual labels 'aids in performance evaluation' refers to evaluation-only label assignment for padded samples using the unchanged longer modality's real labels, not to using ground-truth labels as training inputs. The paper does note a limitation that anchors are selected from complete data and performance declines at very low alignment rates, which is a stated scope restriction rather than a circular step. Any concerns about the unvalidated class-fidelity assumption of Gaussian interpolation or the unspecified kernel width sigma are correctness risks, not circularity: they do not make the reported superiority reduce to the method's own definitions. Accordingly, no specific circular step satisfying the quoted-reduction requirement was found.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claim rests on two heuristic bets: that the random-walk/greedy anchor selection covers all classes, and that Gaussian interpolation between the most different pairs creates class-preserving virtual samples. Neither is proven. The second is the load-bearing assumption of the consistency-aware padding module. The many unspecified hyperparameters (alpha, d, na, sigma, a, m) add further free choices that affect the reported results.

free parameters (6)
  • alpha (self-repellent walk strength) = 0.5
    Set by hand in Section 3.1 to balance mixing and efficiency; no sensitivity analysis is provided.
  • random walk steps and step size (nw, wl) = lookup by N in Eq. (3)
    Chosen via a heuristic table based on dataset size; no justification for the specific values.
  • greedy radius d = not specified
    The radius used to mark covered nodes in Eq. (7) is never given a value in the paper.
  • number of anchors na = not specified
    The paper mentions selecting a required number of anchor points but does not state how this number is chosen; na is only defined as the size of the deduplicated index union.
  • margin m and coefficient a in noise-contrastive loss = not specified
    Both appear in Eq. (13) and control the loss behavior, but their values or selection method are not reported.
  • Gaussian kernel width sigma = not specified
    The kernel width in Eq. (20) is a free parameter, but no value or tuning procedure is given.
assumptions (4)
  • domain assumption The cosine-similarity transition matrix P (Eq. 4) captures semantic structure suitable for anchor selection.
    The method builds P from cosine similarities and uses it for random walks; no proof is given that the resulting anchors preserve cluster information.
  • ad hoc to paper Self-repellent random walk combined with the greedy algorithm selects anchors covering every class.
    Section 3.1 asserts this property, but the greedy radius d and anchor count are unspecified, and no coverage guarantee is derived.
  • ad hoc to paper Gaussian-kernel interpolation between the most dissimilar aligned pairs produces virtual samples that preserve the class structure.
    This is the core assumption of the consistency-aware padding module (Section 3.3, Eq. 19). The paper does not justify it theoretically or empirically beyond aggregate clustering metrics.
  • standard math Standard algorithms (Hungarian alignment, k-means, Gaussian kernels) behave correctly on the re-represented and padded data.
    The method relies on well-known algorithmic building blocks without re-deriving them.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Consistency-Aware Padding for Incomplete Multi-Modal Alignment Clustering Based on Self-Repellent Greedy Anchor Search." pith.science (2026). https://pith.science/paper/Q65J7UDH

@misc{pith2026250703917,
  author       = {Pith},
  title        = {Pith review of: Consistency-Aware Padding for Incomplete Multi-Modal Alignment Clustering Based on Self-Repellent Greedy Anchor Search},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Q65J7UDH}},
  note         = {Machine review of arXiv:2507.03917}
}
read the original abstract

Multimodal representation is faithful and highly effective in describing real-world data samples' characteristics by describing their complementary information. However, the collected data often exhibits incomplete and misaligned characteristics due to factors such as inconsistent sensor frequencies and device malfunctions. Existing research has not effectively addressed the issue of filling missing data in scenarios where multiview data are both imbalanced and misaligned. Instead, it relies on class-level alignment of the available data. Thus, it results in some data samples not being well-matched, thereby affecting the quality of data fusion. In this paper, we propose the Consistency-Aware Padding for Incomplete Multimodal Alignment Clustering Based on Self-Repellent Greedy Anchor Search(CAPIMAC) to tackle the problem of filling imbalanced and misaligned data in multimodal datasets. Specifically, we propose a self-repellent greedy anchor search module(SRGASM), which employs a self-repellent random walk combined with a greedy algorithm to identify anchor points for re-representing incomplete and misaligned multimodal data. Subsequently, based on noise-contrastive learning, we design a consistency-aware padding module (CAPM) to effectively interpolate and align imbalanced and misaligned data, thereby improving the quality of multimodal data fusion. Experimental results demonstrate the superiority of our method over benchmark datasets. The code will be publicly released at https://github.com/Autism-mm/CAPIMAC.git.

Figures

Figures reproduced from arXiv: 2507.03917 by the authors.

Figure 1
Figure 1. Incomplete partial alignment example graph. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Incomplete and unaligned CAPIMAC model. In the Data Pre-processing, we handle incomplete and misaligned multimodal data. In [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Flowchart of the Consistency-aware Padding Module: The [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

36 extracted references · 33 canonical work pages

  1. [1]

    Node embedding accelerates randoms walk on a graph

    [Aung and Ohsaki, 2024] Han Nay Aung and Hiroyuki Ohsaki. Node embedding accelerates randoms walk on a graph. In 2024 IEEE 48th Annual Computers, Soft- ware, and Applications Conference (COMPSAC) , pages 537–545. IEEE,

    work page 2024

  2. [8]

    Partially view-aligned clustering

    [Huang et al., 2020] Zhenyu Huang, Peng Hu, Joey Tianyi Zhou, Jiancheng Lv, and Xi Peng. Partially view-aligned clustering. Advances in Neural Information Processing Systems, 33:2892–2902,

    work page 2020

  3. [10]

    Large- scale multi-view subspace clustering in linear time

    [Kang et al., 2020] Zhao Kang, Wangtao Zhou, Zhitong Zhao, Junming Shao, Meng Han, and Zenglin Xu. Large- scale multi-view subspace clustering in linear time. InPro- ceedings of the AAAI conference on artificial intelligence, volume 34, pages 4412–4419,

    work page 2020

  4. [11]

    Node embedding for homophilous graphs with argew: Augmentation of random walks by graph edge weights

    [Kim et al., 2023] Jun Hee Kim, Jaeman Son, Hyunsoo Kim, and Eunjo Lee. Node embedding for homophilous graphs with argew: Augmentation of random walks by graph edge weights. arXiv preprint arXiv:2308.05957,

    arXiv 2023

  5. [12]

    Incomplete multi- view clustering via prototype-based imputation

    [Li et al., 2023b] Haobin Li, Yunfan Li, Mouxing Yang, Peng Hu, Dezhong Peng, and Xi Peng. Incomplete multi- view clustering via prototype-based imputation. arXiv preprint arXiv:2301.11045,

    arXiv

  6. [13]

    From text to pixels: a context-aware semantic synergy solu- tion for infrared and visible image fusion

    [Li et al., 2023d] Xingyuan Li, Yang Zou, Jinyuan Liu, Zhiying Jiang, Long Ma, Xin Fan, and Risheng Liu. From text to pixels: a context-aware semantic synergy solu- tion for infrared and visible image fusion. arXiv preprint arXiv:2401.00421,

    arXiv

  7. [14]

    Large graph construction for scalable semi-supervised learning

    [Liu et al., 2010] Wei Liu, Junfeng He, and Shih-Fu Chang. Large graph construction for scalable semi-supervised learning. In Proceedings of the 27th international con- ference on machine learning (ICML-10) , pages 679–686. Citeseer,

    work page 2010

  8. [16]

    Fuzzy granular anomaly detection using markov random walk

    [Liu et al., 2023] Chang Liu, Zhong Yuan, Baiyang Chen, Hongmei Chen, and Dezhong Peng. Fuzzy granular anomaly detection using markov random walk. Informa- tion Sciences, 646:119400,

    work page 2023

Show all 36 references
  1. [17]

    Dual consensus anchor learning for fast multi-view clustering

    [Qin et al., 2024] Yalan Qin, Chuan Qin, Xinpeng Zhang, and Guorui Feng. Dual consensus anchor learning for fast multi-view clustering. IEEE Transactions on Image Pro- cessing,

    2024

  2. [18]

    Adap- tive clustering algorithm based on knn and density.Pattern Recognition Letters, 104:37–44,

    [Shi et al., 2018] Bing Shi, Lixin Han, and Hong Yan. Adap- tive clustering algorithm based on knn and density.Pattern Recognition Letters, 104:37–44,

    2018

  3. [20]

    A novel multi-label classifi- cation algorithm based on k-nearest neighbor and random walk

    [Wang et al., 2020] Zhen-Wu Wang, Si-Kai Wang, Ben-Ting Wan, and William Wei Song. A novel multi-label classifi- cation algorithm based on k-nearest neighbor and random walk. International Journal of Distributed Sensor Net- works, 16(3):1550147720911892,

    2020

  4. [21]

    Skywalker: Efficient alias-method-based graph sampling and random walk on gpus

    [Wang et al., 2021] Pengyu Wang, Chao Li, Jing Wang, Taolei Wang, Lu Zhang, Jingwen Leng, Quan Chen, and Minyi Guo. Skywalker: Efficient alias-method-based graph sampling and random walk on gpus. In 2021 30th International Conference on Parallel Architectures and Compilation T...

    2021

  5. [22]

    Contrastive consensus graph learning for multi-view clustering

    [Wang et al., 2022] Shiping Wang, Xincan Lin, Zihan Fang, Shide Du, and Guobao Xiao. Contrastive consensus graph learning for multi-view clustering. IEEE/CAA Journal of Automatica Sinica, 9(11):2027–2030,

    2022

  6. [23]

    Fualign: Cross-lingual entity alignment via multi-view representation learning of fused knowledge graphs

    [Wang et al., 2023] Chenxu Wang, Zhenhao Huang, Yue Wan, Junyu Wei, Junzhou Zhao, and Pinghui Wang. Fualign: Cross-lingual entity alignment via multi-view representation learning of fused knowledge graphs. Infor- mation Fusion, 89:41–52,

    2023

  7. [24]

    Scalable and structural multi- view graph clustering with adaptive anchor fusion

    [Wang et al., 2024] Siwei Wang, Xinwang Liu, Suyuan Liu, Wenxuan Tu, and En Zhu. Scalable and structural multi- view graph clustering with adaptive anchor fusion. IEEE Transactions on Image Processing,

    2024

  8. [25]

    Deep in- complete multi-view clustering via mining cluster comple- mentarity

    [Xu et al., 2022] Jie Xu, Chao Li, Yazhou Ren, Liang Peng, Yujie Mo, Xiaoshuang Shi, and Xiaofeng Zhu. Deep in- complete multi-view clustering via mining cluster comple- mentarity. In Proceedings of the AAAI conference on arti- ficial intelligence, volume 36, pages 8761–8769,

    2022

  9. [26]

    Partially view- aligned representation learning with noise-robust con- trastive loss

    [Yang et al., 2021] Mouxing Yang, Yunfan Li, Zhenyu Huang, Zitao Liu, Peng Hu, and Xi Peng. Partially view- aligned representation learning with noise-robust con- trastive loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 1134– 1143,

    2021

  10. [27]

    Robust multi- view clustering with incomplete information

    [Yang et al., 2022] Mouxing Yang, Yunfan Li, Peng Hu, Jin- feng Bai, Jiancheng Lv, and Xi Peng. Robust multi- view clustering with incomplete information. IEEE Trans- actions on Pattern Analysis and Machine Intelligence , 45(1):1055–1069,

    2022

  11. [28]

    Cross-view graph matching for incomplete multi-view clustering

    [Yang et al., 2023] Jing-Hua Yang, Le-Le Fu, Chuan Chen, Hong-Ning Dai, and Zibin Zheng. Cross-view graph matching for incomplete multi-view clustering. Neuro- computing, 515:79–88,

    2023

  12. [29]

    Geometric-inspired graph-based incomplete multi-view clustering

    [Yang et al., 2024] Zequn Yang, Han Zhang, Yake Wei, Zheng Wang, Feiping Nie, and Di Hu. Geometric-inspired graph-based incomplete multi-view clustering. Pattern Recognition, 147:110082,

    2024

  13. [30]

    Consensus kernel k-means clustering for in- complete multiview data

    [Ye et al., 2017] Yongkai Ye, Xinwang Liu, Qiang Liu, and Jianping Yin. Consensus kernel k-means clustering for in- complete multiview data. Computational intelligence and neuroscience, 2017(1):3961718,

    2017

  14. [31]

    Dvsai: Diverse view-shared anchors based incomplete multi-view clustering

    [Yu et al., 2024] Shengju Yu, Siwei Wang, Pei Zhang, Miao Wang, Ziming Wang, Zhe Liu, Liming Fang, En Zhu, and Xinwang Liu. Dvsai: Diverse view-shared anchors based incomplete multi-view clustering. In Proceedings of the AAAI Conference on Artificial Intelligence , volume 38, ...

    2024

  15. [32]

    Seman- tic invariant multi-view clustering with fully incomplete information

    [Zeng et al., 2023] Pengxin Zeng, Mouxing Yang, Yiding Lu, Changqing Zhang, Peng Hu, and Xi Peng. Seman- tic invariant multi-view clustering with fully incomplete information. IEEE Transactions on Pattern Analysis and Machine Intelligence,

    2023

  16. [33]

    An end-to-end framework for partial view-aligned clustering with graph structure

    [Zhao et al., 2023b] Liang Zhao, Qiongjie Xie, Sontao Wu, and Shubin Ma. An end-to-end framework for partial view-aligned clustering with graph structure. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. IEEE,

    2023

  17. [34]

    Anchor based multi-view clustering for partially view-aligned data

    [Zhao et al., 2024c] Liang Zhao, Yukun Yuan, Qiongjie Xie, and Ziyue Wang. Anchor based multi-view clustering for partially view-aligned data. In 2024 IEEE International Conference on Multimedia and Expo (ICME) , pages 1–5. IEEE,

    2024

  18. [35]

    A survey and an empiri- cal evaluation of multi-view clustering approaches

    [Zhou et al., 2024] Lihua Zhou, Guowang Du, Kevin L ¨u, Lizheng Wang, and Jingwei Du. A survey and an empiri- cal evaluation of multi-view clustering approaches. ACM Computing Surveys, 56(7):1–38,

    2024

  19. [36]

    Incomplete multi-view clustering with partially mapped instances and clusters

    [Zong et al., 2021] Linlin Zong, Faqiang Miao, Xianchao Zhang, Xinyue Liu, and Hong Yu. Incomplete multi-view clustering with partially mapped instances and clusters. Knowledge-Based Systems, 212:106615, 2021

    2021

  20. [2010]

    Efficient and effective regularized incomplete multi-view clustering

    [Liu et al., 2020] Xinwang Liu, Miaomiao Li, Chang Tang, Jingyuan Xia, Jian Xiong, Li Liu, Marius Kloft, and En Zhu. Efficient and effective regularized incomplete multi-view clustering. IEEE transactions on pattern anal- ysis and machine intelligence, 43(8):2634–2646,

    2020

  21. [2016]

    Robust variational contrastive learning for partially view-unaligned clustering

    [He et al., 2024] Changhao He, Hongyuan Zhu, Peng Hu, and Xi Peng. Robust variational contrastive learning for partially view-unaligned clustering. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 4167–4176,

    2024

  22. [2017]

    Self- repellent random walks on general graphs-achieving min- imal sampling variance via nonlinear markov chains

    [Doshi et al., 2023] Vishwaraj Doshi, Jie Hu, et al. Self- repellent random walks on general graphs-achieving min- imal sampling variance via nonlinear markov chains. In International Conference on Machine Learning , pages 8403–8423. PMLR,

    2023

  23. [2018]

    Multi-view subspace clustering via adaptive graph learn- ing and late fusion alignment

    [Tang et al., 2023] Chuan Tang, Kun Sun, Chang Tang, Xiao Zheng, Xinwang Liu, Jun-Jie Huang, and Wei Zhang. Multi-view subspace clustering via adaptive graph learn- ing and late fusion alignment. Neural Networks, 165:333– 343,

    2023

  24. [2020]

    Unified view imputation and feature selection learning for incomplete multi-view data

    [Huang et al., 2024] Yanyong Huang, Zongxin Shen, Tianrui Li, and Fengmao Lv. Unified view imputation and feature selection learning for incomplete multi-view data. arXiv preprint arXiv:2401.10549,

    2024 arXiv

  25. [2021]

    metapath2vec: Scalable representa- tion learning for heterogeneous networks

    [Dong et al., 2017] Yuxiao Dong, Nitesh V Chawla, and Ananthram Swami. metapath2vec: Scalable representa- tion learning for heterogeneous networks. In Proceed- ings of the 23rd ACM SIGKDD international conference on knowledge discovery and data mining, pages 135–144,

    2017

  26. [2022]

    Temporal multi-view graph convolutional networks for citywide traffic volume infer- ence

    [Dai et al., 2021] Shaojie Dai, Jinshuai Wang, Chao Huang, Yanwei Yu, and Junyu Dong. Temporal multi-view graph convolutional networks for citywide traffic volume infer- ence. In 2021 IEEE International Conference on Data Mining (ICDM), pages 1042–1047. IEEE,

    2021

  27. [2023]

    node2vec: Scalable feature learning for networks

    [Grover and Leskovec, 2016] Aditya Grover and Jure Leskovec. node2vec: Scalable feature learning for networks. In Proceedings of the 22nd ACM SIGKDD international conference on Knowledge discovery and data mining, pages 855–864,

    2016

  28. [2024]

    Graph generation and diffusion using random walks

    [Cai et al., 2022] Wenyu Cai, Gilbert Chen Ye, and Hao Zhou. Graph generation and diffusion using random walks. Highlights in Science, Engineering and Technol- ogy, 16:490–494,

    2022

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.