Pith. sign in

REVIEW 3 major objections 6 minor 107 references

Learning Binarized Representations with Pseudo-positive Sample Enhancement for Efficient Graph Collaborative Filtering

T0 review · 3 major / 6 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 By training binarized user-item embeddings with pseudo-positive samples—items the full-precision teacher ranks highly plus synthesized hard negatives—BiGeaR++ recovers 98–108% of full-precision recommendation accuracy while running…

desk verdict BiGeaR++ is a credible incremental extension of BiGeaR with real efficiency gains, but the pseudo-positive distillation mechanism is under-validated and the ablations need significance tests. read the letter →

arxiv 2506.02750 v1 pith:OD3YES3I submitted 2025-06-03 cs.IR

Yankai Chen , Yue Que , Xinni Zhang , Chen Ma , Irwin King This is my paper
classification cs.IR
keywords RecommendersystemsEmbeddingbinarizationPseudo-positivesamplesGraphconvolutionalnetworksKnowledgedistillationCollaborativefilteringEfficientinference
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

This paper claims that the accuracy gap between full-precision and binarized graph collaborative filtering can be largely closed by treating 'pseudo-positive' samples as extra supervisory signals. The proposed model, BiGeaR++, adds two mechanisms to its predecessor BiGeaR: dual inference distillation, in which the binary student learns to rank not only items the user interacted with but also non-interacted items the full-precision teacher scores highly, and pseudo-positive embedding sample synthesis, which generates hard-to-distinguish latent negatives by mixing positive item embeddings into sampled negatives and selecting the hardest per layer. Across five datasets, these additions improve Recall@20 and NDCG@20 by 1.08–10.26% over BiGeaR, reaching 98.11–108.26% of the best full-precision model's capability. The practical payoff is that near-full-precision recommendations could be served from binary embeddings with about 8x faster inference and 9x memory compression, making CPU-based deployment more realistic.

What carries the argument

The central object is the pseudo-positive sample, defined in two ways: (1) real non-interacted items that the full-precision teacher embeddings rank in the top R for a user, used as ranking targets in the second distillation loss $\mathcal{L}_{\text{ID2}}$; and (2) synthetic latent embeddings built by mixing a positive item's embeddings into randomly sampled negatives and selecting the hardest candidate per layer by inner-product score. These samples carry the argument by providing the binary student with informative ranking targets beyond the sparse ground-truth positives, and the paper's ablations show that removing these components degrades Recall@20 by up to about 10% on Amazon-Book. The layer-wise quantization with per-layer scalers and the Dirac-delta gradient estimator are supporting machinery that protects the signal these samples are meant to teach.

What would settle it

Take the top-R teacher-selected pseudo-positive items for each user and measure how many appear in that user's held-out test interactions; if this precision is at or below that of random non-interacted items, or if replacing the teacher's top-R with randomly chosen items leaves Recall@20 unchanged, then the $\mathcal{L}_{\text{ID2}}$ term's gains come from something other than accurate pseudo-labels.

Watch

Extended reading notes

Core claim

The discovery is that the information loss in embedding binarization is not a single event but a series of losses at distinct stages—feature expressivity, ranking inference, and gradient estimation—and that each can be reduced with a targeted mechanism. BiGeaR++ quantizes embeddings layer by layer with per-layer scalers, distills ranking knowledge from full-precision to binarized embeddings in a layer-wise fashion using both ground-truth positives and pseudo-positive items selected by the teacher, synthesizes pseudo-positive embeddings via positive mix-up and hard sample selection in both the full-precision and binarized spaces, and estimates gradients through an approximation of the Dirac delta function instead of the straight-through estimator. The paper reports that the two new pseudo-positive components yield consistent gains of about 1%–10% over BiGeaR across MovieLens, Gowalla, Pinterest, Yelp2018, and Amazon-Book, with the ablations showing that removing the pseudo-positive distillation term or the synthesized-sample training hurts more than removing the positive-item distillation term.

Load-bearing premise

The load-bearing assumption is that items a full-precision teacher ranks highest among a user's non-interacted set are likely enough to be true positives that training the binary student to rank them highly improves held-out recommendation; nothing in the paper measures how often these pseudo-labels are actually correct.

Editorial extensions

If this is right

  • Online recommendation can be served from binary embeddings at roughly 8x faster inference and 9x lower memory, with quality at 98–108% of the best full-precision model.
  • The two new components integrate with the existing BiGeaR modules without retuning, since the gains hold across five datasets with different sparsity levels.
  • The layer-wise design means accuracy can be adjusted through embedding dimension and layer count rather than by abandoning binarization.
  • The ablation results imply that pseudo-positive supervision, not merely larger embedding size, drives the improvement over prior binary methods.

Reading between the lines

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

  • A testable implication left implicit by the paper is that the pseudo-positive distillation should help most on sparse datasets where ground-truth positives are rare; the reported largest gains on Amazon-Book are consistent with this, but the paper does not analyze the correlation.
  • One could validate pseudo-label quality directly by measuring how often teacher-selected top-R items appear in held-out interactions; if precision is low on sparse datasets, tuning R per dataset may be needed.
  • The synthesis approach resembles hard-negative mining from contrastive learning, which suggests BiGeaR++ could be combined with theoretical negative-sampling guidance, such as sampling negatives from the estimated positive distribution, to sharpen the student further.
  • Because the teacher and student share the same graph convolution backbone, an implicit extension is to distill a larger pre-trained teacher into a smaller binary student, which the current architecture does not explore.
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

3 major / 6 minor

Summary. This paper presents BiGeaR++, an extension of the binarized graph collaborative filtering model BiGeaR. The main additions are a dual inference distillation objective (L_ID1 over interacted items and L_ID2 over teacher-selected pseudo-positive non-interacted items, Eqs. 11 and 13) and a pseudo-positive embedding sample synthesizer that mixes positive and negative embeddings and selects hard samples (Section 2.4). These are combined with BiGeaR's layer-wise quantization and Dirac-based gradient estimator. Experiments on MovieLens, Gowalla, Pinterest, Yelp2018, and Amazon-Book report consistent Recall@20/NDCG@20 gains over BiGeaR (1.08%-10.26%), near or above LightGCN, about 8x faster inference and 9x lower memory on MovieLens. The paper also provides complexity analyses and makes code available.

Significance. If the claims hold, BiGeaR++ is a practically relevant contribution: it demonstrates that a 1-bit recommender can operate within a few percent of full-precision models while replacing floating-point inner products with XNOR/Popcount operations, a result with direct deployment value. The strengths include the explicit bitwise decomposition in Eq. 7, the reproducible hyperparameter table, held-out evaluation with external metrics, and a public implementation. The main weaknesses are the unvalidated pseudo-positive labels in L_ID2 and the absence of significance testing for several small ablation differences; these need to be addressed before the component-level attributions can be accepted.

major comments (3)
  1. [Section 2.3.2, Eq. (13), Table 8] The L_ID2 mechanism is built on the assumption that the teacher's top-R non-interacted items are pseudo-positives whose high ranking by the student should be encouraged. The manuscript explicitly concedes that these items are 'not necessarily all ground-truth positives,' but it never measures their precision. In implicit-feedback CF, the teacher's highest-scoring unseen items often overlap with popular items, so the observed gains (w/o ID2 drops of 0.43%-3.10% in Table 8) may reflect exposure to popular items rather than transfer of genuine preference knowledge. Please report the held-out positive rate among selected S_tch(u) items versus random non-interacted items on each dataset and include a control variant of L_ID2 that uses randomly chosen non-interacted items as targets.
  2. [Section 4.2, Tables 8-9] The paper states that improvements are 'statistically significant' but reports no significance tests. Several ablation differences are very small (e.g., -0.05% to -0.38% Recall@20 for w/o ID1 in Table 8, and a '-' entry for Pinterest NDCG), while the reported standard deviations differ substantially between BiGeaR++ (0.01-0.03) and BiGeaR (0.11-0.33). Paired tests across seeds (e.g., bootstrap or paired t-test) are needed for the BiGeaR++ versus BiGeaR comparison and for each component ablation; without them, the component-wise attribution in RQ3 is not established.
  3. [Section 3.2, Table 2] The training-complexity table is not currently verifiable because several entries are malformed or use notation inconsistently (e.g., 'O(2SdE2L/B)', and N is used both for the number of items and for the average number of interacted items in the same row). Since the paper presents resource efficiency as a core contribution and uses these asymptotics to compare with LightGCN, the table should be restated with fully defined symbols and corrected formulas.
minor comments (6)
  1. [Table 6 and Section 4.2] The 'Capability' percentages are computed against the underlined best full-precision model (DGCF on Pinterest), while the text says the comparison is with LightGCN; please state the reference model for each capability row.
  2. [Tables 8-9] The '-' entries in the difference rows are unexplained; please either report the measured values or state that the difference is below the reported precision.
  3. [Algorithm 1] The 'if with inference distillation' branch appears to define the binarized training phase and the 'else' branch the full-precision pre-training phase, but the condition name is not defined in the pseudocode; rename for clarity.
  4. [Section 2.4] The synthesized samples in Eqs. (15) and (18) are used as negatives in BPR losses (Eqs. 21-22) but are called 'pseudo-positive samples'; please align terminology with their role in the loss, since L_ID2's real item samples are also called pseudo-positive.
  5. [Eq. (16)] With c=1, beta can take the endpoint values 0 and 1, which would fully suppress either the negative or the positive component in Eq. (15); state whether such cases occur or are clipped.
  6. [Throughout] There are several typographical issues (e.g., 'recommedner' in Section 6, 'wavelines' in the Table 6 caption); a copyedit pass is needed.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the reported gains are measured against held-out Recall/NDCG benchmarks, not against quantities fitted or defined by the model.

full rationale

The paper's central claims are empirical: BiGeaR++ improves Recall@20 and NDCG@20 over its predecessor BiGeaR and approaches full-precision LightGCN (Tables 6 and 7). These numbers are computed on held-out test interactions, which are external to the training objective in Eq. (23). The dual inference distillation losses (Eqs. 11 and 13) select pseudo-positive items using the full-precision teacher's own scores, and the sample synthesizer (Eqs. 15-18) constructs hard negatives; these are training regularizers rather than the quantities being evaluated. The feature uniqueness analysis in Eqs. (26)-(28) is a mathematical consequence of the chosen symmetric normalization in Eq. (1), not an assumption of the performance claim. The paper's use of the authors' earlier BiGeaR [9] as the architectural base and as a baseline is a normal extension, and the gain over BiGeaR is measured externally rather than imported from the self-citation. The Table 2 footnote stating that BiGeaR implemented an early version of inference distillation is a novelty concern, not evidence that the new results are determined by their inputs. Likewise, the unvalidated precision of the pseudo-positive labels in L_ID2 is a label-noise and correctness risk, not a circularity: the final metrics remain external. No load-bearing step reduces, by construction, to its own inputs.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The method rests on standard recommender assumptions and a set of per-dataset hyperparameters. No new physical entities are introduced; 'pseudo-positive samples' are a training construct built from existing item embeddings and teacher predictions.

free parameters (7)
  • J (number of negative candidates in sample synthesis) = 8, 4, 20, 20, 2 per dataset
    Tuned per dataset in Table 5; Figure 7 shows the optimal J varies by dataset.
  • c (mixup control factor) = 1.0 on four datasets, 0.01 on Amazon-Book
    Hand-set per dataset to control the degree of positive mixing in Eq. 16; reported in Table 5.
  • R (number of pseudo-positive real samples for L_ID2) = 50 on MovieLens, 25 on the other four datasets
    Tuned hyperparameter, reported in Table 5.
  • lambda_2 (sharpness of ranking-aware weights) = 0.1
    Fixes the exponential decay of w_k in Eq. 12.
  • gamma (gradient approximation sharpness) = 1
    Controls the Gaussian approximation of the Dirac delta in Eq. 25.
  • lambda (L2 regularization coefficient) = 1e-4, 5e-5, 1e-4, 1e-4, 1e-6 per dataset
    Tuned over {1e-6, 1e-5, 1e-4, 1e-3} in Table 5.
  • eta (learning rate) = 1e-3 or 5e-4
    Tuned per dataset over {1e-4, 1e-3, 1e-2}, reported in Table 5.
assumptions (5)
  • domain assumption Non-interacted items are generally negative, and the highest-scoring non-interacted items under the full-precision teacher are a useful set of pseudo-positives for distillation.
    L_ID2 in Section 2.3.2 treats top-R non-interacted items as positive-ranking targets for the student; the paper acknowledges they are not guaranteed positives and uses w_k to reweight them.
  • standard math LightGCN-style symmetric-normalized graph convolution without nonlinearity is a suitable backbone for binarized collaborative filtering.
    The paper adopts Eq. 1 from LightGCN as the propagation operator for both teacher and student embeddings.
  • domain assumption The hardest mixup negatives, selected by highest inner product with the user embedding, carry the most informative ranking signal.
    Motivated by prior negative-sampling theory; the paper does not independently verify this distributional assumption on each dataset.
  • standard math The derivative of sign(phi) behaves as a Dirac delta, approximated by a zero-centered Gaussian with sharpness gamma.
    Mathematical identity from Section 2.5 used to justify the gradient estimator in Eq. 25.
  • standard math The L1-norm embedding scaler alpha = (1/d) ||v||_1 faithfully scalarizes each binarized segment.
    Eq. 3 uses the mean absolute value as the scalar for sign(v); this is the least-squares optimal scalar for a fixed sign pattern and underpins both prediction and the acceleration formula.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning Binarized Representations with Pseudo-positive Sample Enhancement for Efficient Graph Collaborative Filtering." pith.science (2026). https://pith.science/paper/OD3YES3I

@misc{pith2026250602750,
  author       = {Pith},
  title        = {Pith review of: Learning Binarized Representations with Pseudo-positive Sample Enhancement for Efficient Graph Collaborative Filtering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OD3YES3I}},
  note         = {Machine review of arXiv:2506.02750}
}
read the original abstract

Learning vectorized embeddings is fundamental to many recommender systems for user-item matching. To enable efficient online inference, representation binarization, which embeds latent features into compact binary sequences, has recently shown significant promise in optimizing both memory usage and computational overhead. However, existing approaches primarily focus on numerical quantization, neglecting the associated information loss, which often results in noticeable performance degradation. To address these issues, we study the problem of graph representation binarization for efficient collaborative filtering. Our findings indicate that explicitly mitigating information loss at various stages of embedding binarization has a significant positive impact on performance. Building on these insights, we propose an enhanced framework, BiGeaR++, which specifically leverages supervisory signals from pseudo-positive samples, incorporating both real item data and latent embedding samples. Compared to its predecessor BiGeaR, BiGeaR++ introduces a fine-grained inference distillation mechanism and an effective embedding sample synthesis approach. Empirical evaluations across five real-world datasets demonstrate that the new designs in BiGeaR++ work seamlessly well with other modules, delivering substantial improvements of around 1%-10% over BiGeaR and thus achieving state-of-the-art performance compared to the competing methods. Our implementation is available at https://github.com/QueYork/BiGeaR-SS.

Figures

Figures reproduced from arXiv: 2506.02750 by the authors.

Figure 1
Figure 1. Illustration of BiGeaR++. users and items, transforming continuous full-precision representations into discrete binarized versions. These binarized representations enable significant reductions in model size and speed up inference, leveraging low-bit arithmetic on devices where CPUs are generally more cost-effective than high-end GPUs [2, 3]. Despite its promising potential, simply stacking CF methods with binarizat… view at source ↗
Figure 2
Figure 2. BiGeaR++ first pre-trains the full-precision embeddings and then triggers the (1) graph layer-wise quantization, (2) dual inference distillation, (3) pseudo-positive sample synthesis, and (4) accurate gradient estimation to learn the binarized representations (Best view in color). Preliminaries: graph convolution. The general approach is to learn node representations by iteratively propagating and aggregating latent… view at source ↗
Figure 3
Figure 3. Acceleration with bitwise operations. Instead of making 𝛼 (𝑙) 𝑢 and 𝛼 (𝑙) 𝑖 learnable, this deterministic computation approach is simple yet effective in providing the scaling functionality while significantly reducing the parameter search space. After 𝐿 layers of quantization and scaling, we construct the following binarized embedding table for each graph node 𝑥 as: A𝑥 = {𝛼 (0) 𝑥 , 𝛼 (1) 𝑥 , · · · , 𝛼 (𝐿) 𝑥 }, Q𝑥 =… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Illustration of our pseudo-positive sample synthesis approach. 2.4.1 Pseudo-positive Sample Synthesis. Instead of relying on real items for negative sampling [48, 49], and inspired by [33, 85], we develop a pseudo-positive sample synthesizer to generate highly informat…
Figure 5
Figure 5. Figure 5: Illustration of our gradient estimation approach. [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: Study of graph layer-wise quantization on BiGeaR++ [PITH_FULL_IMAGE:figures/full_fig_p019_6.png]
Figure 7
Figure 7. Figure 7: Varying 𝐽 of negative items in sample synthesis. 4.6.3 Size 𝐽 of Negative Items in Sample Synthesis. To validate the effect of 𝐽 settings in pseudo-positive sample synthesis, we conduct experiments across all datasets by varying 𝐽. The results of Recall@20 and NDCG@20 …
Figure 8
Figure 8. Figure 8: Gradient estimator comparison w.r.t. Recall@20 and NDCG@20 [PITH_FULL_IMAGE:figures/full_fig_p022_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

107 extracted references · 70 canonical work pages

  1. [1]

    Rohan Anil, Gabriel Pereyra, Alexandre Passos, Robert Ormandi, George E Dahl, and Geoffrey E Hinton. 2018. Large scale distributed neural network training through online distillation. ICLR

    2018

  2. [2]

    Mehdi Bahri, Gaétan Bahl, and Stefanos Zafeiriou. 2021. Binary Graph Neural Networks. In CVPR. 9492–9501

    2021

  3. [3]

    Ron Banner, Itay Hubara, Elad Hoffer, and Daniel Soudry. 2018. Scalable methods for 8-bit training of neural networks. NeurIPS 31

    2018

  4. [4]

    Yoshua Bengio, Nicholas Léonard, and Aaron Courville. 2013. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv

    2013

  5. [5]

    Rianne van den Berg, Thomas N Kipf, and Max Welling. 2017. Graph convolutional matrix completion. arXiv preprint arXiv:1706.02263

    arXiv 2017

  6. [6]

    Ronald Newbold Bracewell and Ronald N Bracewell. 1986. The Fourier transform and its applications. Vol. 31999. McGraw-Hill New York. Manuscript submitted to ACM BiGeaR++: Learning Binarized Representations with Pseudo-positive Sample Learning Enhancement 25

    1986

  7. [7]

    Zhangjie Cao, Mingsheng Long, Jianmin Wang, and Philip S Yu. 2017. Hashnet: Deep learning to hash by continuation. In ICCV. 5608–5617

    2017

  8. [8]

    Jingyuan Chen, Hanwang Zhang, Xiangnan He, Liqiang Nie, Wei Liu, and Tat-Seng Chua. 2017. Attentive collaborative filtering: Multimedia recommendation with item-and component-level attention. In SIGIR. 335–344

    2017

Show all 107 references
  1. [9]

    Yankai Chen, Huifeng Guo, Yingxue Zhang, Chen Ma, Ruiming Tang, Jingjie Li, and Irwin King. 2022. Learning binarized graph representations with multi-faceted quantization reinforcement for top-k recommendation. In SIGKDD. 168–178

    2022

  2. [10]

    Yankai Chen, Quoc-Tuan Truong, Xin Shen, Jin Li, and Irwin King. 2024. Shopping trajectory representation learning with pre-training for e-commerce customer understanding and recommendation. In SIGKDD. 385–396

    2024

  3. [11]

    Yankai Chen, Quoc-Tuan Truong, Xin Shen, Ming Wang, Jin Li, Jim Chan, and Irwin King. 2023. Topological Representation Learning for E-commerce Shopping Behaviors. (2023)

    2023

  4. [12]

    Yankai Chen, Taotao Wang, Yixiang Fang, and Yunyu Xiao. 2025. Semi-supervised node importance estimation with informative distribution modeling for uncertainty regularization. In WWW. 3108–3118

    2025

  5. [13]

    Yankai Chen, Menglin Yang, Yingxue Zhang, Mengchen Zhao, Ziqiao Meng, Jianye Hao, and Irwin King. 2022. Modeling Scale-free Graphs with Hyperbolic Geometry for Knowledge-aware Recommendation. WSDM

    2022

  6. [14]

    Yankai Chen, Yaming Yang, Yujing Wang, Jing Bai, Xiangchen Song, and Irwin King. 2022. Attentive Knowledge-aware Graph Convolutional Networks with Collaborative Guidance for Personalized Recommendation. ICDE

    2022

  7. [15]

    Yankai Chen, Yifei Zhang, Huifeng Guo, Ruiming Tang, and Irwin King. 2022. An effective post-training embedding binarization approach for fast online top-k passage matching. In AACL. 102–108

    2022

  8. [16]

    Yankai Chen, Yifei Zhang, Yingxue Zhang, Huifeng Guo, Jingjie Li, Ruiming Tang, Xiuqiang He, and Irwin King. 2021. Towards low-loss 1-bit quantization of user-item representations for top-k recommendation. arXiv preprint arXiv:2112.01944 (2021)

    2021 arXiv

  9. [17]

    Paul Covington, Jay Adams, and Emre Sargin. 2016. Deep neural networks for youtube recommendations. In Recsys. 191–198

    2016

  10. [18]

    Sajad Darabi, Mouloud Belbahri, Matthieu Courbariaux, and Vahid Partovi Nia. 2018. Bnn+: Improved binary network training. arXiv

    2018

  11. [19]

    Xiang Deng and Zhongfei Zhang. 2021. Graph-free knowledge distillation for graph neural networks. IJCAI (2021)

    2021

  12. [20]

    Kaituo Feng, Changsheng Li, Ye Yuan, and Guoren Wang. 2022. Freekd: Free-direction knowledge distillation for graph neural networks. In SIGKDD. 357–366

    2022

  13. [21]

    Step function. 2022. https://en.wikipedia.org/wiki/Heaviside_step_function

    2022

  14. [22]

    Xue Geng, Hanwang Zhang, Jingwen Bian, and Tat-Seng Chua. 2015. Learning image and user features for recommendation in social networks. In ICCV

    2015

  15. [23]

    Aristides Gionis, Piotr Indyk, Rajeev Motwani, et al. 1999. Similarity search in high dimensions via hashing. In VLDB, Vol. 99. 518–529

    1999

  16. [24]

    Ruihao Gong, Xianglong Liu, Shenghu Jiang, Tianxiang Li, Peng Hu, Jiazhen Lin, Fengwei Yu, and Junjie Yan. 2019. Differentiable soft quantization: Bridging full-precision and low-bit neural networks. In ICCV. 4852–4861

    2019

  17. [25]

    William L Hamilton, Rex Ying, and Jure Leskovec. 2017. Inductive representation learning on large graphs. In NeurIPS. 1025–1035

    2017

  18. [26]

    Ruining He and Julian McAuley. 2016. Modeling the visual evolution of fashion trends with one-class collaborative filtering. In WWW. 507–517

    2016

  19. [27]

    Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yongdong Zhang, and Meng Wang. 2020. Lightgcn: Simplifying and powering graph convolution network for recommendation. In SIGIR. 639–648

    2020

  20. [28]

    Xiangnan He, Zhankui He, Jingkuan Song, Zhenguang Liu, Yu-Gang Jiang, and Tat-Seng Chua. 2018. Nais: Neural attentive item similarity model for recommendation. TKDE 30, 12, 2354–2366

    2018

  21. [29]

    Xiangnan He, Lizi Liao, Hanwang Zhang, Liqiang Nie, Xia Hu, and Tat-Seng Chua. 2017. Neural collaborative filtering. In WWW. 173–182

    2017

  22. [30]

    Xiangnan He, Hanwang Zhang, Min-Yen Kan, and Tat-Seng Chua. 2016. Fast matrix factorization for recommendation with implicit feedback. In SIGIR

    2016

  23. [31]

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. 2015. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531

    2015 arXiv

  24. [32]

    Hengtong Hu, Lingxi Xie, Richang Hong, and Qi Tian. 2020. Creating something from nothing: Unsupervised knowledge distillation for cross-modal hashing. In CVPR. 3123–3132

    2020

  25. [33]

    Tinglin Huang, Yuxiao Dong, Ming Ding, Zhen Yang, Wenzheng Feng, Xinyu Wang, and Jie Tang. 2021. MixGCF: An Improved Training Method for Graph Neural Network-based Recommender Systems. In KDD

    2021

  26. [34]

    Eric Jang, Shixiang Gu, and Ben Poole. 2017. Categorical reparameterization with gumbel-softmax. In 5th ICLR

    2017

  27. [35]

    Young Kyun Jang, Geonmo Gu, Byungsoo Ko, Isaac Kang, and Nam Ik Cho. 2022. Deep hash distillation for image retrieval. InECCV. Springer, 354–371

    2022

  28. [36]

    Chaitanya K Joshi, Fayao Liu, Xu Xun, Jie Lin, and Chuan Sheng Foo. 2022. On representation knowledge distillation for graph neural networks. TNNLS (2022)

    2022

  29. [37]

    Wang-Cheng Kang, Derek Zhiyuan Cheng, Tiansheng Yao, Xinyang Yi, Ting Chen, Lichan Hong, and Ed H Chi. 2021. Learning to embed categorical features without embedding tables for recommendation. SIGKDD

    2021

  30. [38]

    Wang-Cheng Kang and Julian McAuley. 2019. Candidate generation with binary codes for large-scale top-n recommendation. InCIKM. 1523–1532

    2019

  31. [39]

    Diederik P Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In ICLR

    2015

  32. [40]

    Thomas N Kipf and Max Welling. 2017. Semi-supervised classification with graph convolutional networks. In 5th ICLR

    2017

  33. [41]

    Pang Wei Koh and Percy Liang. 2017. Understanding black-box predictions via influence functions. In ICML. PMLR, 1885–1894

    2017

  34. [42]

    Yehuda Koren, Robert Bell, and Chris Volinsky. 2009. Matrix factorization techniques for recommender systems. Computer 42, 8, 30–37. Manuscript submitted to ACM 26 Yankai Chen et al

    2009

  35. [43]

    Guohao Li, Matthias Muller, Ali Thabet, and Bernard Ghanem. 2019. Deepgcns: Can gcns go as deep as cnns?. In ICCV. 9267–9276

    2019

  36. [44]

    Hao Li, Wei Liu, and Heng Ji. 2014. Two-Stage Hashing for Fast Document Retrieval.. In ACL (2). 495–500

    2014

  37. [45]

    Qimai Li, Zhichao Han, and Xiao-Ming Wu. 2018. Deeper insights into graph convolutional networks for semi-supervised learning. In AAAI

    2018

  38. [46]

    Yangning Li, Yinghui Li, Xinyu Wang, Yong Jiang, Zhen Zhang, Xinran Zheng, Hui Wang, Hai-Tao Zheng, Fei Huang, Jingren Zhou, et al

  39. [47]

    Yang Li, Kangbo Liu, Ranjan Satapathy, Suhang Wang, and Erik Cambria. 2024. Recent developments in recommender systems: A survey. IEEE Computational Intelligence Magazine 19, 2 (2024), 78–95

    2024

  40. [48]

    Yangning Li, Tingwei Lu, Hai-Tao Zheng, Yinghui Li, Shulin Huang, Tianyu Yu, Jun Yuan, and Rui Zhang. 2024. MESED: A multi-modal entity set expansion dataset with fine-grained semantic classes and hard negative entities. In Proceedings of the AAAI Conference on Artificial Inte...

    2024

  41. [49]

    Yangning Li, Qingsong Lv, Tianyu Yu, Yinghui Li, Xuming Hu, Wenhao Jiang, Hai-Tao Zheng, and Hui Wang. 2025. UltraWiki: Ultra-Fine-Grained Entity Set Expansion with Negative Seed Entities . In 2025 IEEE 41st International Conference on Data Engineering (ICDE)

    2025

  42. [50]

    Yang Li, Suhang Wang, Quan Pan, Haiyun Peng, Tao Yang, and Erik Cambria. 2019. Learning binary codes with neural collaborative filtering for efficient recommendation systems. KBS 172 (2019), 64–75

    2019

  43. [51]

    Dawen Liang, Laurent Charlin, James McInerney, and David M Blei. 2016. Modeling user exposure in recommendation. In WWW. 951–961

    2016

  44. [52]

    Zijing Liang, Yanjie Xu, Yifan Hong, Penghui Shang, Qi Wang, Qiang Fu, and Ke Liu. 2024. A Survey of Multimodel Large Language Models. In Proceedings of the 3rd International Conference on Computer, Artificial Intelligence and Control Engineering. 405–409

    2024

  45. [53]

    Sida Lin, Zhouyi Zhang, Yankai Chen, Chenhao Ma, Yixiang Fang, Shan Dai, and Guangli Lu. 2024. Effective Job-market Mobility Prediction with Attentive Heterogeneous Knowledge Learning and Synergy. In CIKM. 3897–3901

    2024

  46. [54]

    Xiaofan Lin, Cong Zhao, and Wei Pan. 2017. Towards accurate binary convolutional neural network. In NeurIPS

    2017

  47. [55]

    Chunlei Liu, Wenrui Ding, Xin Xia, Yuan Hu, Baochang Zhang, Jianzhuang Liu, Bohan Zhuang, and Guodong Guo. 2019. RBCN: Rectified binary convolutional networks for enhancing the performance of 1-bit DCNNs. arXiv

    2019

  48. [56]

    Kang Liu, Zhenhua Huang, Chang-Dong Wang, Beibei Gao, and Yunwen Chen. 2024. Fine-grained learning behavior-oriented knowledge distillation for graph neural networks. TNNLS (2024)

    2024

  49. [57]

    Xianglong Liu, Junfeng He, Cheng Deng, and Bo Lang. 2014. Collaborative hashing. In CVPR. 2139–2146

    2014

  50. [58]

    Fangyuan Luo, Yankai Chen, Jun Wu, and Yidong Li. 2025. Rank Gap Sensitive Deep AUC maximization for CTR prediction. Pattern Recognition (2025), 111496

    2025

  51. [59]

    Chris J Maddison, Andriy Mnih, and Yee Whye Teh. 2017. The concrete distribution: A continuous relaxation of discrete random variables. In 5th ICLR

    2017

  52. [60]

    Yoon-Joo Park and Alexander Tuzhilin. 2008. The long tail of recommender systems and how to leverage it. In RecSys. 11–18

    2008

  53. [61]

    Haotong Qin, Ruihao Gong, Xianglong Liu, Mingzhu Shen, Ziran Wei, Fengwei Yu, and Jingkuan Song. 2020. Forward and backward information retention for accurate binary neural networks. In CVPR. 2250–2259

    2020

  54. [62]

    Zexuan Qiu, Jiahong Liu, Yankai Chen, and Irwin King. 2024. Hihpq: Hierarchical hyperbolic product quantization for unsupervised image retrieval. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 38. 4614–4622

    2024

  55. [63]

    Zexuan Qiu, Jieming Zhu, Yankai Chen, Guohao Cai, Weiwen Liu, Zhenhua Dong, and Irwin King. 2024. EASE: Learning Lightweight Semantic Feature Adapters from Large Language Models for CTR Prediction. In CIKM. 4819–4827

    2024

  56. [64]

    Mohammad Rastegari, Vicente Ordonez, Joseph Redmon, and Ali Farhadi. 2016. Xnor-net: Imagenet classification using binary convolutional neural networks. In ECCV. Springer, 525–542

    2016

  57. [65]

    Steffen Rendle and Christoph Freudenthaler. 2014. Improving pairwise learning for item recommendation from implicit feedback. In CIKM. 273–282

    2014

  58. [66]

    Steffen Rendle, Christoph Freudenthaler, Zeno Gantner, and Lars Schmidt-Thieme. 2012. BPR: Bayesian personalized ranking from implicit feedback. arXiv

    2012

  59. [67]

    Mingyue Su, Guanghua Gu, Xianlong Ren, Hao Fu, and Yao Zhao. 2021. Semi-supervised knowledge distillation for cross-modal hashing. IEEE Transactions on Multimedia 25 (2021), 662–675

    2021

  60. [68]

    Qiaoyu Tan, Ninghao Liu, Xing Zhao, Hongxia Yang, Jingren Zhou, and Xia Hu. 2020. Learning to hash with GNNs for recommender systems. In WWW. 1988–1998

    2020

  61. [69]

    Wentao Tan, Lei Zhu, Jingjing Li, Huaxiang Zhang, and Junwei Han. 2022. Teacher-student learning: Efficient hierarchical message aggregation hashing for cross-modal retrieval. IEEE Transactions on Multimedia 25 (2022), 4520–4532

    2022

  62. [70]

    Jiaxi Tang and Ke Wang. 2018. Learning compact ranking models with high performance for recommender system. In SIGKDD. 2289–2298

    2018

  63. [71]

    Yijun Tian, Shichao Pei, Xiangliang Zhang, Chuxu Zhang, and Nitesh Chawla. 2023. Knowledge distillation on graphs: A survey. Comput. Surveys (2023)

    2023

  64. [72]

    Petar Veličković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. 2018. Graph attention networks. ICLR

    2018

  65. [73]

    Chenxu Wang, Yue Wan, Zhenhao Huang, Panpan Meng, and Pinghui Wang. 2022. AEP: Aligning knowledge graphs via embedding propagation. Neurocomputing 507 (2022), 130–144. Manuscript submitted to ACM BiGeaR++: Learning Binarized Representations with Pseudo-positive Sample Learning...

    2022

  66. [74]

    Chenxu Wang, Zhizhong Wan, Panpan Meng, Shihao Wang, and Zhanggong Wang. 2024. Graph contrastive learning with high-order feature interactions and adversarial Wasserstein-distance-based alignment. IJMLC (2024), 1–12

    2024

  67. [75]

    Chenxu Wang, Zhiyang Zhu, Panpan Meng, and Yumo Qiu. 2022. Leveraging network structure for efficient dynamic negative sampling in network embedding. Information Sciences 606 (2022), 853–863

    2022

  68. [76]

    Fangxin Wang, Kay Liu, Sourav Medya, and Philip S. Yu. 2025. BANGS: Game-theoretic Node Selection for Graph Self-Training. In ICLR

    2025

  69. [77]

    Fangxin Wang, Yuqing Liu, Kay Liu, Yibo Wang, Sourav Medya, and Philip S. Yu. 2024. Uncertainty in Graph Neural Networks: A Survey. TMLR (2024)

    2024

  70. [78]

    Junfu Wang, Yunhong Wang, Zhen Yang, Liang Yang, and Yuanfang Guo. 2021. Bi-gcn: Binary graph convolutional network. InCVPR. 1561–1570

    2021

  71. [79]

    Jingdong Wang, Ting Zhang, Nicu Sebe, Heng Tao Shen, et al. 2017. A survey on learning to hash. TPAMI 40, 4, 769–790

    2017

  72. [80]

    Xiang Wang, Xiangnan He, Meng Wang, Fuli Feng, and Tat-Seng Chua. 2019. Neural graph collaborative filtering. In SIGIR. 165–174

    2019

  73. [81]

    Xiang Wang, Hongye Jin, An Zhang, Xiangnan He, Tong Xu, and Tat-Seng Chua. 2020. Disentangled graph collaborative filtering. In SIGIR. 1001–1010

    2020

  74. [82]

    Wei Wei, Xubin Ren, Jiabin Tang, Qinyong Wang, Lixin Su, Suqi Cheng, Junfeng Wang, Dawei Yin, and Chao Huang. 2024. Llmrec: Large language models with graph augmentation for recommendation. InProceedings of the 17th ACM International Conference on Web Search and Data Mining. 806–815

    2024

  75. [83]

    Felix Wu, Amauri Souza, Tianyi Zhang, Christopher Fifty, Tao Yu, and Kilian Weinberger. 2019. Simplifying graph convolutional networks. In ICML. PMLR

    2019

  76. [84]

    Lirong Wu, Haitao Lin, Yufei Huang, and Stan Z Li. 2022. Knowledge distillation improves graph structure augmentation for graph neural networks. NeurIPS 35 (2022), 11815–11827

    2022

  77. [85]

    Xi Wu, Liangwei Yang, Jibing Gong, Chao Zhou, Tianyu Lin, Xiaolong Liu, and Philip S. Yu. 2023. Dimension Independent Mixup for Hard Negative Sample in Collaborative Filtering. In CIKM (CIKM ’23). ACM. https://doi.org/10.1145/3583780.3614845

    2023

  78. [86]

    Yaozu Wu, Yankai Chen, Zhishuai Yin, Weiping Ding, and Irwin King. 2023. A survey on graph embedding techniques for biomedical data: Methods and applications. Information Fusion 100 (2023), 101909

    2023

  79. [87]

    Zonghan Wu, Shirui Pan, Fengwen Chen, Guodong Long, Chengqi Zhang, and S Yu Philip. 2020. A comprehensive survey on graph neural networks. IEEE TNNLS 32, 1, 4–24

    2020

  80. [88]

    Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. 2020. Self-training with noisy student improves imagenet classification. In CVPR. 10687–10698

    2020

  81. [89]

    Keyulu Xu, Chengtao Li, Yonglong Tian, Tomohiro Sonobe, Ken-ichi Kawarabayashi, and Stefanie Jegelka. 2018. Representation learning on graphs with jumping knowledge networks. In ICML. PMLR, 5453–5462

    2018

  82. [90]

    Jiwei Yang, Xu Shen, Jun Xing, Xinmei Tian, Houqiang Li, Bing Deng, Jianqiang Huang, and Xian-sheng Hua. 2019. Quantization networks. In CVPR. 7308–7316

    2019

  83. [91]

    Menglin Yang, Min Zhou, Jiahong Liu, Defu Lian, and Irwin King. 2022. HRCF: Enhancing collaborative filtering via hyperbolic geometric regularization. In WebConf. 2462–2471

    2022

  84. [92]

    Yiding Yang, Jiayan Qiu, Mingli Song, Dacheng Tao, and Xinchao Wang. 2020. Distilling knowledge from graph convolutional networks. In CVPR. 7074–7083

    2020

  85. [93]

    Zhen Yang, Ming Ding, Chang Zhou, Hongxia Yang, Jingren Zhou, and Jie Tang. 2020. Understanding negative sampling in graph representation learning. In SIGKDD. 1666–1676

    2020

  86. [94]

    Rex Ying, Ruining He, Kaifeng Chen, Pong Eksombatchai, William L Hamilton, and Jure Leskovec. 2018. Graph convolutional neural networks for web-scale recommender systems. In SIGKDD. 974–983

    2018

  87. [95]

    Yang Yu, Meiyu Liang, Mengran Yin, Kangkang Lu, Junping Du, and Zhe Xue. 2024. Unsupervised Multimodal Graph Contrastive Semantic Anchor Space Dynamic Knowledge Distillation Network for Cross-Media Hash Retrieval. In ICDE. IEEE, 4699–4708

    2024

  88. [96]

    Hanwang Zhang, Fumin Shen, Wei Liu, Xiangnan He, Huanbo Luan, and Tat-Seng Chua. 2016. Discrete collaborative filtering. In SIGIR. 325–334

    2016

  89. [97]

    Jifan Zhang, Fangxin Wang, Philip S Yu, Kaize Ding, and Shixiang Zhu. 2025. Topology-Aware Conformal Prediction for Stream Networks. arXiv preprint arXiv:2503.04981 (2025)

    2025

  90. [98]

    Weinan Zhang, Tianqi Chen, Jun Wang, and Yong Yu. 2013. Optimizing top-n collaborative filtering via dynamic negative item sampling. In SIGIR. 785–788

    2013

  91. [99]

    Xinni Zhang, Yankai Chen, Cuiyun Gao, Qing Liao, Shenglin Zhao, and Irwin King. 2022. Knowledge-aware neural networks with personalized feature referencing for cold-start recommendation. arXiv preprint arXiv:2209.13973 (2022)

    2022 arXiv

  92. [100]

    Xinni Zhang, Yankai Chen, Chenhao Ma, Yixiang Fang, and Irwin King. 2024. Influential Exemplar Replay for Incremental Learning in Recommender Systems. In AAAI, Vol. 38. 9368–9376

    2024

  93. [101]

    Yifei Zhang, Yankai Chen, Zixing Song, and Irwin King. 2023. Contrastive cross-scale graph knowledge synergy. In SIGKDD. 3422–3433

    2023

  94. [102]

    Yan Zhang, Defu Lian, and Guowu Yang. 2017. Discrete personalized ranking for fast collaborative filtering from implicit feedback. In AAAI, Vol. 31

    2017

  95. [103]

    Yifei Zhang, Hao Zhu, Yankai Chen, Zixing Song, Piotr Koniusz, Irwin King, et al. 2023. Mitigating the popularity bias of graph collaborative filtering: A dimensional collapse perspective. Advances in Neural Information Processing Systems 36 (2023), 67533–67550. Manuscript sub...

    2023

  96. [104]

    Yifei Zhang, Hao Zhu, Zixing Song, Yankai Chen, Xinyu Fu, Ziqiao Meng, Piotr Koniusz, and Irwin King. 2024. Geometric view of soft decorrelation in self-supervised learning. In SIGKDD. 4338–4349

    2024

  97. [105]

    Yuhan Zhao, Rui Chen, Riwei Lai, Qilong Han, Hongtao Song, and Li Chen. 2023. Augmented Negative Sampling for Collaborative Filtering. In RecSys. 256–266

    2023

  98. [106]

    Sheng Zhou, Yucheng Wang, Defang Chen, Jiawei Chen, Xin Wang, Can Wang, and Jiajun Bu. 2021. Distilling holistic knowledge with graph neural networks. In ICCV. 10387–10396. Manuscript submitted to ACM

    2021

  99. [2024]

    arXiv preprint arXiv:2411.02937 (2024)

    Benchmarking multimodal retrieval augmented generation with dynamic vqa dataset and self-adaptive planning agent. arXiv preprint arXiv:2411.02937 (2024)

    2024 arXiv

Pith tools

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