Pith. sign in

REVIEW 4 major objections 4 minor 42 references

MixDec Sampling: A Soft Link-based Sampling Method of Graph Neural Network for Recommendation

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

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

Pith's one-line read This paper proposes MixDec Sampling, a soft-link sampling method for GNN-based recommenders that claims consistent recommendation gains across three backbones and three benchmarks.

desk verdict Plausible soft-link sampling for GNN recommenders, but the evaluation only beats uniform sampling and doesn't support the 'significantly and consistently' claim. read the letter →

arxiv 2502.08161 v1 pith:IWPFYHCL submitted 2025-02-12 cs.IR cs.AI

Xiangjin Xie , Yuxin Chen , Ruipeng Wang , Kai Ouyang , Zihan Zhang , Hai-Tao Zheng , Buyue Qian , Hansen Zheng
show 3 more authors
Bo Hu Chengxiang Zhuo Zang Li
This is my paper · ORCID
classification cs.IRcs.AI
keywords graphneuralnetworkrecommendersystemnegativesamplingsoftlinkMixupdataaugmentationBFSdecaycollaborativefiltering
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

MixDec Sampling is a plug-in training method for graph-neural-network recommenders that replaces the binary view of user-item pairs (positive or negative) with soft links whose weights encode graded proximity. It has two modules: Mixup Sampling synthesizes new nodes and soft links by linearly interpolating between sampled positive and negative items, and Decay Sampling attaches soft links to multi-hop neighbors with weights that decay according to BFS path counts. The authors claim that this consistently improves recommendation quality across three benchmarks and three GNN backbones, reporting average MRR gains of 18.3% for GraphSAGE and 21.1% for GCN. The intended contribution is a general sampling mechanism that any GNN-based recommender can adopt without changing its architecture.

What carries the argument

The central mechanism is the soft link weight used as a regression target for the model's similarity score. Mixup Sampling creates synthetic nodes via $e_s = \lambda e_i + (1-\lambda) e_j$ and link weights via $w_s = \lambda w_i + (1-\lambda) w_j$; Decay Sampling assigns multi-hop neighbors weights $w_d = \rho + (1-\rho) r_d / r_{\max}$ from BFS path counts. Both auxiliary losses are mean absolute error between $\sigma(e_v^\top e_s)$ (or $\sigma(e_v^\top e_d)$) and the corresponding soft weight, turning sampling from a binary classification signal into a graded proximity signal.

What would settle it

Run the same GNN backbones on the same datasets with the same synthetic and multi-hop pairs but with random soft weights in [0,1] replacing the MixDec weights; if random weights produce comparable MRR and Hit@30 gains, the specific proximity semantics of the soft links are not what drives the improvement.

Watch

Extended reading notes

Core claim

The paper claims that modeling sampling relationships as soft links, instead of hard positive/negative pairs, produces better node embeddings for recommendation. Mixup Sampling generates synthetic items by mixing feature vectors of positive-positive or positive-negative pairs drawn for an anchored user, with weight $w_s = \lambda w_i + (1-\lambda) w_j$ and $\lambda \sim \mathrm{Beta}(\alpha,\beta)$, then regresses the sigmoid of the user-synthetic-item dot product toward $w_s$. Decay Sampling builds soft links from the anchored user to l-hop neighbors with weight $w_d = \rho + (1-\rho) r_d / r_{\max}$, where $r_d$ is the number of BFS paths reaching the neighbor, and regresses the sigmoid dot product toward $w_d$. The joint loss is $L = L_{ns} + L_m + L_d$, and the experiments show that this joint objective improves MRR and Hit@30 over uniform negative sampling on Amazon-Book, Yelp2018, and Last-FM with GraphSAGE, GCN, and GAT.

Load-bearing premise

The load-bearing premise is that the hand-defined soft weights, namely Mixup interpolation weights and BFS decay weights, are sensible supervision targets for the model's similarity scores; if they do not track true item relevance, the auxiliary losses could distort the learned embeddings.

Editorial extensions

If this is right

  • Existing GNN recommenders can adopt MixDec Sampling as a drop-in replacement for the sampling and auxiliary loss, with no change to the aggregation architecture.
  • Nodes with few neighbors receive synthetic soft-link training pairs, which should improve recommendation for sparse and long-tail items.
  • Multi-hop structure is fed into training through decay-weighted links, so the model can digest graph reachability beyond one-hop edges.
  • The added sampling overhead is bounded: MixDec training takes no more than roughly 30 percent longer than uniform negative sampling per run.

Reading between the lines

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

  • Because the method is model-agnostic, a natural next test is on newer GNN recommenders beyond GCN, GraphSAGE, and GAT; the paper's own experiments do not show whether those gains transfer.
  • The soft-weight regression could improve performance partly by regularizing the embedding space rather than by encoding true proximity; a random-weight control would separate those effects.
  • The random-walk approximation for decay weights is proposed but not evaluated; testing whether approximate weights preserve the reported gains on large graphs would tell whether the method scales.
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

4 major / 4 minor

Summary. The paper proposes MixDec Sampling, a soft-link-based sampling augmentation for GNN-based recommender systems. It combines a Mixup Sampling module, which synthesizes new nodes by linearly interpolating features and link weights of positive/negative samples, and a Decay Sampling module, which assigns soft link weights to multi-hop neighbors according to BFS path counts. The combined auxiliary losses (Eqs. (13) and (14)) are added to the standard negative-sampling loss. The authors evaluate the method with GraphSAGE, GCN, and GAT on Amazon-Book, Yelp2018, and Last-FM, reporting MRR and Hit@30 improvements over Uniform Negative Sampling, with average MRR gains of 18.3% for GraphSAGE and 21.1% for GCN claimed in the abstract.

Significance. If the reported gains hold up under rigorous evaluation, MixDec Sampling would be a simple, plug-in augmentation that could improve several GNN-based recommenders, which is practically useful. The paper contributes a reasonable heuristic idea—replacing hard binary pair relations with continuously weighted soft links derived from graph structure—and includes pseudocode, released source code, an ablation study, and a complexity discussion. However, the significance is currently limited by the empirical evaluation: no error bars or significance tests, comparisons only against a single weak baseline, and internal inconsistencies in the 'consistently better' claim. The soft-weight regression targets are also not empirically validated, leaving the mechanism's benefit unestablished.

major comments (4)
  1. [§IV-B, Table II] The central claim that MixDec sampling 'significantly and consistently improve[s]' recommendation performance is not supported by the evidence presented. Section IV-A4 states that training runs for 2,500 epochs but does not report multiple seeds, standard deviations, confidence intervals, or significance tests, so 'significantly' is unsubstantiated. Moreover, Table II contains direct counterexamples to the 'consistently' claim: on Last-FM, Decay Sampling alone outperforms MixDec for GCN on MRR (0.1892 vs 0.1840) and Hit@30 (0.5642 vs 0.5596), and for GraphSAGE on MRR (0.2353 vs 0.2280); on Yelp2018, Decay Sampling also edges out MixDec for GCN on MRR (0.1842 vs 0.1822). The authors should report repeated-run statistics, run significance tests, and either temper the claim or demonstrate that MixDec is reliably superior to its own components.
  2. [§IV-A2 and §V-B] The only negative-sampling baseline in the experiments is Uniform Negative Sampling, despite Section V-B discussing stronger model-based methods such as DNS, IRGAN, MixGCF, and SRNS. Because the paper's contribution is explicitly a sampling method, the improvement over uniform sampling alone is insufficient to establish practical value. At least one state-of-the-art sampler (e.g., MixGCF, which is also Mixup-related) should be compared under the same backbone, training schedule, and evaluation protocol; otherwise, the claimed 'improvement' may simply reflect the weakness of the baseline.
  3. [§IV-C and §IV-A4] The hyperparameter study (Tables V and VI) tunes α, β, ρ, and k on Amazon-Book with GCN, and the implementation details do not report per-dataset or per-model hyperparameter settings for the main experiments. If the Amazon-Book GCN-tuned values are used for all datasets and all backbones, the evaluation risks overfitting the augmentation schedule to one dataset/architecture, and the reported gains on other datasets may not reflect good practice. The authors should either tune these parameters on each dataset's validation split and describe the chosen values, or provide a sensitivity analysis demonstrating that the results are stable across a reasonable hyperparameter range.
  4. [§III-D, Eqs. (13) and (14)] The auxiliary losses regress the sigmoid of user-item embedding dot products to hand-designed soft weights w_s and w_d from Eqs. (4) and (8). The paper does not provide any evidence that these regression targets are well-calibrated proxies for relevance, and if the soft weights systematically conflict with the true interaction signal, the auxiliary losses could distort the embedding space. This concern is secondary to the missing statistical validation, but it is load-bearing for why the method helps: the authors should at least show that the gains are robust to reasonable monotone transformations of the decay weights and that the mixup weights are not simply acting as label noise.
minor comments (4)
  1. [§III-B, random-walk paragraph] The text says 'Then, rd in (10) is the number of occurrences of sampled node d in the walk sequences', but Equation (10) is the Decay Sampling loss; the definition of rd was given in Equation (8). Please correct the cross-reference.
  2. [Fig. 4 caption] The caption contains a typo: 'BFS-based dacay weight' should be 'BFS-based decay weight'.
  3. [Fig. 1 caption] The caption says 'MixDec Samping', a typo for 'MixDec Sampling'.
  4. [§IV-B1] The bullet states that Uniform Negative Sampling 'performs the worst in all evaluations', but Table II shows that Mixup Sampling on Last-FM with GAT yields MRR 0.1820, which is worse than Uniform Negative Sampling's 0.1834. This statement should be qualified.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: MixDec's soft-link target weights are defined from graph structure and sampled labels, independent of the evaluated metrics.

full rationale

The paper does not offer a formal derivation from assumptions to predictions; MixDec Sampling is an empirical heuristic. The soft-link weights used as auxiliary training targets are constructed directly from the input graph: Eq. (4) linearly mixes sampled positive/negative link weights with a Beta draw, and Eq. (8) sets decay weights from BFS/random-walk path counts. Losses (13) and (14) regress sigmoid dot-product scores to those fixed graph-derived weights. None of these definitions references MRR, Hit@K, or any held-out target, so the reported gains in Table II are not forced by construction. The ablation and parameter studies compare the same training procedure with and without the added terms under a shared aggregator and loss implementation. No load-bearing self-citation is used to justify the central claim, and no uniqueness theorem is imported to rule out alternatives. The absence of significance tests and strong negative-sampling baselines is an empirical-evidence weakness, not circularity.

Assumptions & free parameters 7 free parameters · 5 assumptions · 1 invented entities

The method relies on several hand-designed components: the mixup weight interpolation, the BFS path-count decay formula, and four hyperparameters chosen by validation. No external evidence or theoretical guarantee is provided for these design choices.

free parameters (7)
  • alpha (Beta distribution shape) = 0.2
    Tuned on Amazon-Book GCN; best MRR in Table V. Used for mixup interpolation.
  • beta (Beta distribution shape) = 0.5
    Tuned on Amazon-Book GCN; best MRR in Table V. Used for mixup interpolation.
  • rho (decay weight lower bound) = 0.5
    Best value in Table VI; maps decay weights to [rho, 1] in Eq. (8).
  • k (top-k decay nodes) = 500
    Best value in Table VI; number of decay items retained per anchor.
  • cm (number of mixup samples) = 5
    Stated in Section IV-A4.
  • cd (number of decay samples) = 5
    Stated in Section IV-A4.
  • l (number of aggregation layers / BFS depth) = 3
    Section IV-A2: three-layer structure; also determines BFS hop count.
assumptions (5)
  • standard math Mixup linear interpolation yields valid training examples with soft labels (Zhang et al. 2018).
    Relied upon for the Mixup Sampling module, Eq. (3)-(4).
  • standard math BFS path counts enumerate all reachable paths up to l hops.
    Used to define decay weights r_d in Eq. (8).
  • domain assumption Regressing sigmoid(e_u^T e_v) to hand-defined soft weights w_s and w_d improves recommendation accuracy.
    The entire training signal of MixDec depends on this unproven assumption; no calibration or theoretical justification given.
  • domain assumption Path-count-based decay weights capture item-user proximity that is useful for recommendation.
    Eq. (8) assumes more paths imply stronger relevance; this is a heuristic without empirical support.
  • ad hoc to paper The decay weight formula w_d = rho + (1-rho) r_d / r_max is appropriate without derivation.
    Introduced in Section III-B, Eq. (8), with no derivation or external basis.
invented entities (1)
  • Soft link (weighted training edge with weight in [0,1])
    purpose: Supervision target for synthetic (mixup) and decay (BFS-hop) items against the anchor user embedding.
    The soft link is a training construct; it is not added to the graph for message passing and does not correspond to a measured quantity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MixDec Sampling: A Soft Link-based Sampling Method of Graph Neural Network for Recommendation." pith.science (2026). https://pith.science/paper/IWPFYHCL

@misc{pith2026250208161,
  author       = {Pith},
  title        = {Pith review of: MixDec Sampling: A Soft Link-based Sampling Method of Graph Neural Network for Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IWPFYHCL}},
  note         = {Machine review of arXiv:2502.08161}
}
read the original abstract

Graph neural networks have been widely used in recent recommender systems, where negative sampling plays an important role. Existing negative sampling methods restrict the relationship between nodes as either hard positive pairs or hard negative pairs. This leads to the loss of structural information, and lacks the mechanism to generate positive pairs for nodes with few neighbors. To overcome limitations, we propose a novel soft link-based sampling method, namely MixDec Sampling, which consists of Mixup Sampling module and Decay Sampling module. The Mixup Sampling augments node features by synthesizing new nodes and soft links, which provides sufficient number of samples for nodes with few neighbors. The Decay Sampling strengthens the digestion of graph structure information by generating soft links for node embedding learning. To the best of our knowledge, we are the first to model sampling relationships between nodes by soft links in GNN-based recommender systems. Extensive experiments demonstrate that the proposed MixDec Sampling can significantly and consistently improve the recommendation performance of several representative GNN-based models on various recommendation benchmarks.

Figures

Figures reproduced from arXiv: 2502.08161 by the authors.

Figure 1
Figure 1. An overview of the MixDec sampling. MixDec consists of Mixup Sampling and Decay Sampling. 1) Based on sampled positive and negative items [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. An illustration of the Mixup sampling. User [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. An example of BFS-based decay weights of soft links. For the [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: An illustration of the MixDec sampling. For example, in the figure, [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Distributions of Different Beta(α, β). The x-axis represents the value of λ, and the y-axis represents the corresponding probability density. TABLE V IMPACT OF BETA DISTRIBUTION α \ β 0.2 0.5 0.8 1 5 8 0.2 0.1442 0.1570 0.1520 0.1498 0.1419 0.1472 0.5 0.1455 0.1445 0.1…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

42 extracted references · 34 canonical work pages

  1. [1]

    Graph neural networks in recommender systems: a survey,

    S. Wu, F. Sun, W. Zhang, X. Xie, and B. Cui, “Graph neural networks in recommender systems: a survey,” ACM Computing Surveys (CSUR) , 2020

    work page 2020

  2. [2]

    Semi-supervised classification with graph convolutional networks,

    T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” ICLR’17, 2017

    work page 2017

  3. [3]

    Inductive representation learning on large graphs,

    W. Hamilton, Z. Ying, and J. Leskovec, “Inductive representation learning on large graphs,” in NIPS’17, pp. 1024–1034, 2017

    work page 2017

  4. [4]

    Graph attention networks,

    P. Veli ˇckovi´c, G. Cucurull, A. Casanova, A. Romero, P. Lio, and Y . Ben- gio, “Graph attention networks,” arXiv preprint arXiv:1710.10903, 2017

    arXiv 2017

  5. [5]

    Graph convolutional neural networks for web-scale rec- ommender systems,

    R. Ying, R. He, K. Chen, P. Eksombatchai, W. L. Hamilton, and J. Leskovec, “Graph convolutional neural networks for web-scale rec- ommender systems,” in Proceedings of the 24th ACM SIGKDD interna- tional conference on knowledge discovery & data mining , pp. 974–983, 2018

    work page 2018

  6. [6]

    Lightgcn: Simplifying and powering graph convolution network for recommenda- tion,

    X. He, K. Deng, X. Wang, Y . Li, Y . Zhang, and M. Wang, “Lightgcn: Simplifying and powering graph convolution network for recommenda- tion,” in Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval , pp. 639–648, 2020

    work page 2020

  7. [7]

    Bpr: Bayesian personalized ranking from implicit feedback,

    S. Rendle, C. Freudenthaler, Z. Gantner, and L. Schmidt-Thieme, “Bpr: Bayesian personalized ranking from implicit feedback,” arXiv preprint arXiv:1205.2618, 2012

    arXiv 2012

  8. [9]

    Irgan: A minimax game for unifying generative and discriminative information retrieval models,

    J. Wang, L. Yu, W. Zhang, Y . Gong, Y . Xu, B. Wang, P. Zhang, and D. Zhang, “Irgan: A minimax game for unifying generative and discriminative information retrieval models,” in Proceedings of the 40th International ACM SIGIR conference on Research and Development in Information Retrieval , pp. 515–524, 2017

    work page 2017

Show all 42 references
  1. [10]

    Adversarial sampling and training for semi- supervised information retrieval,

    D. H. Park and Y . Chang, “Adversarial sampling and training for semi- supervised information retrieval,” in The World Wide Web Conference , pp. 1443–1453, 2019

    2019

  2. [11]

    Kbgan: Adversarial learning for knowledge graph embeddings,

    L. Cai and W. Y . Wang, “Kbgan: Adversarial learning for knowledge graph embeddings,” arXiv preprint arXiv:1711.04071 , 2017

    2017 arXiv

  3. [12]

    mixup: Beyond empirical risk minimization,

    H. Zhang, M. Cisse, Y . N. Dauphin, and D. Lopez-Paz, “mixup: Beyond empirical risk minimization,” in International Conference on Learning Representations, 2018

    2018

  4. [13]

    Distributed representations of words and phrases and their composi- tionality,

    T. Mikolov, I. Sutskever, K. Chen, G. S. Corrado, and J. Dean, “Distributed representations of words and phrases and their composi- tionality,” Advances in neural information processing systems , vol. 26, 2013

    2013

  5. [14]

    Understand- ing negative sampling in graph representation learning,

    Z. Yang, M. Ding, C. Zhou, H. Yang, J. Zhou, and J. Tang, “Understand- ing negative sampling in graph representation learning,” in Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining , pp. 1666–1676, 2020

    2020

  6. [15]

    Efficient estimation of word representations in vector space,

    T. Mikolov, K. Chen, G. S. Corrado, and J. Dean, “Efficient estimation of word representations in vector space,” in ICLR, 2013

    2013

  7. [16]

    Deepwalk: Online learning of social representations,

    B. Perozzi, R. Al-Rfou, and S. Skiena, “Deepwalk: Online learning of social representations,” in Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining , pp. 701–710, 2014

    2014

  8. [17]

    Amazon-book,

    “Amazon-book,” 2015. http://jmcauley.ucsd.edu/data/amazon

    2015

  9. [18]

    Last-fm,

    “Last-fm,” 2015. https://grouplens.org/datasets/hetrec-2011/

    2015

  10. [19]

    https://www.yelp.com/dataset

    yelp, “Yelp,” 2015. https://www.yelp.com/dataset

    2015

  11. [20]

    Kgat: Knowledge graph attention network for recommendation,

    X. Wang, X. He, Y . Cao, M. Liu, and T.-S. Chua, “Kgat: Knowledge graph attention network for recommendation,” inProceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, pp. 950–958, 2019

    2019

  12. [21]

    Mixgcf: An improved training method for graph neural network-based recommender systems,

    T. Huang, Y . Dong, M. Ding, Z. Yang, W. Feng, X. Wang, and J. Tang, “Mixgcf: An improved training method for graph neural network-based recommender systems,” in Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining , pp. 665–674, 2021

    2021

  13. [22]

    Performance of recommender algorithms on top-n recommendation tasks,

    P. Cremonesi, Y . Koren, and R. Turrin, “Performance of recommender algorithms on top-n recommendation tasks,” in Proceedings of the fourth ACM conference on Recommender systems , pp. 39–46, 2010

    2010

  14. [23]

    Scholarly paper recommendation via user’s recent research interests,

    K. Sugiyama and M.-Y . Kan, “Scholarly paper recommendation via user’s recent research interests,” in Proceedings of the 10th annual joint conference on Digital libraries , pp. 29–38, 2010

    2010

  15. [24]

    Pytorch: An imperative style, high-performance deep learning library,

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, et al. , “Pytorch: An imperative style, high-performance deep learning library,” Advances in neural information processing systems , vol. 32, 2019

    2019

  16. [25]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980 , 2014

    2014 arXiv

  17. [26]

    Fastgcn: fast learning with graph convolutional networks via importance sampling,

    J. Chen, T. Ma, and C. Xiao, “Fastgcn: fast learning with graph convolutional networks via importance sampling,” ICLR’18, 2018

    2018

  18. [27]

    Link prediction approach to collaborative filtering,

    H. Chen, X. Li, and Z. Huang, “Link prediction approach to collaborative filtering,” in Proceedings of the 5th ACM/IEEE-CS Joint Conference on Digital Libraries (JCDL’05) , pp. 141–142, IEEE, 2005

    2005

  19. [28]

    Recommendation algorithm based on link prediction and domain knowledge in retail transactions,

    J. Li, L. Zhang, F. Meng, and F. Li, “Recommendation algorithm based on link prediction and domain knowledge in retail transactions,” Procedia Computer Science , vol. 31, pp. 875–881, 2014

    2014

  20. [29]

    Rankclus: integrating clustering with ranking for heterogeneous information net- work analysis,

    Y . Sun, J. Han, P. Zhao, Z. Yin, H. Cheng, and T. Wu, “Rankclus: integrating clustering with ranking for heterogeneous information net- work analysis,” in Proceedings of the 12th international conference on extending database technology: advances in database technology , pp. ...

    2009

  21. [30]

    Ranking-based clustering of heterogeneous information networks with star network schema,

    Y . Sun, Y . Yu, and J. Han, “Ranking-based clustering of heterogeneous information networks with star network schema,” in Proceedings of the 15th ACM SIGKDD international conference on Knowledge discovery and data mining , pp. 797–806, 2009

    2009

  22. [31]

    Pathsim: Meta path- based top-k similarity search in heterogeneous information networks,

    Y . Sun, J. Han, X. Yan, P. S. Yu, and T. Wu, “Pathsim: Meta path- based top-k similarity search in heterogeneous information networks,” Proceedings of the VLDB Endowment , vol. 4, no. 11, pp. 992–1003, 2011

    2011

  23. [32]

    Predicting links in multi- relational and heterogeneous networks,

    Y . Yang, N. Chawla, Y . Sun, and J. Hani, “Predicting links in multi- relational and heterogeneous networks,” in 2012 IEEE 12th international conference on data mining , pp. 755–764, IEEE, 2012

    2012

  24. [33]

    Multi-relational link prediction in heterogeneous information networks,

    D. Davis, R. Lichtenwalter, and N. V . Chawla, “Multi-relational link prediction in heterogeneous information networks,” in 2011 Interna- tional Conference on Advances in Social Networks Analysis and Mining , pp. 281–288, IEEE, 2011

    2011

  25. [34]

    On sampling strategies for neural network-based collaborative filtering,

    T. Chen, Y . Sun, Y . Shi, and L. Hong, “On sampling strategies for neural network-based collaborative filtering,” in Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 767–776, 2017

    2017

  26. [35]

    Improving pairwise learning for item recommendation from implicit feedback,

    S. Rendle and C. Freudenthaler, “Improving pairwise learning for item recommendation from implicit feedback,” in Proceedings of the 7th ACM international conference on Web search and data mining , pp. 273–282, 2014

    2014

  27. [36]

    Optimizing top-n collaborative filtering via dynamic negative item sampling,

    W. Zhang, T. Chen, J. Wang, and Y . Yu, “Optimizing top-n collaborative filtering via dynamic negative item sampling,” inProceedings of the 36th international ACM SIGIR conference on Research and development in information retrieval, pp. 785–788, 2013

    2013

  28. [37]

    Reinforced negative sampling for recommendation with exposure data.,

    J. Ding, Y . Quan, X. He, Y . Li, and D. Jin, “Reinforced negative sampling for recommendation with exposure data.,” in IJCAI, pp. 2230– 2236, 2019

    2019

  29. [38]

    Simplify and robustify negative sampling for implicit collaborative filtering,

    J. Ding, Y . Quan, Q. Yao, Y . Li, and D. Jin, “Simplify and robustify negative sampling for implicit collaborative filtering,” Advances in Neural Information Processing Systems , vol. 33, pp. 1094–1105, 2020

    2020

  30. [39]

    Manifold mixup: Better representations by interpolat- ing hidden states,

    V . Verma, A. Lamb, C. Beckham, A. Najafi, I. Mitliagkas, D. Lopez-Paz, and Y . Bengio, “Manifold mixup: Better representations by interpolat- ing hidden states,” in International Conference on Machine Learning , pp. 6438–6447, PMLR, 2019

    2019

  31. [40]

    Augmenting data with mixup for sentence classification: An empirical study,

    H. Guo, Y . Mao, and R. Zhang, “Augmenting data with mixup for sentence classification: An empirical study,” arXiv preprint arXiv:1905.08941, 2019

    1905 arXiv

  32. [41]

    Global mixup: Eliminating ambiguity with clustering,

    X. Xie, Y . Li, W. Chen, K. Ouyang, L. Jiang, and H. Zheng, “Global mixup: Eliminating ambiguity with clustering,” arXiv preprint arXiv:2206.02734, 2022

    2022 arXiv

  33. [42]

    Graphmixup: Improv- ing class-imbalanced node classification on graphs by self-supervised context prediction,

    L. Wu, H. Lin, Z. Gao, C. Tan, S. Li, et al. , “Graphmixup: Improv- ing class-imbalanced node classification on graphs by self-supervised context prediction,” arXiv preprint arXiv:2106.11133 , 2021

    2021 arXiv

  34. [43]

    ifmixup: Towards intrusion-free graph mixup for graph classification,

    H. Guo and Y . Mao, “ifmixup: Towards intrusion-free graph mixup for graph classification,” arXiv, vol. 2110, 2021

    2021

Pith tools

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