Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Less is More: Information Bottleneck Denoised Multimedia Recommendation

T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Applying information bottleneck denoising at both feature and graph levels removes irrelevant multimedia signals and lifts recommendation accuracy.

desk verdict Real, plausible gains from a two-level HSIC denoiser, but the Information Bottleneck framing overreaches—especially the graph-level module, whose key equality is asserted, not proven. read the letter →

arxiv 2501.12175 v1 pith:HJ7AUXVD submitted 2025-01-21 cs.IR

Yonghui Yang , Le Wu , Zhuangzhuang He , Zhengwei Wu , Richang Hong , Meng Wang This is my paper · ORCID
classification cs.IR
keywords multimediarecommendationinformationbottleneckfeaturedenoisinggraphstructurelearningHilbert-Schmidtindependencecriterionmultimodalcollaborativefiltering
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 argues that multimedia recommenders are held back by task-irrelevant content that generic pre-trained extractors inject into item representations, such as image backgrounds and redundant text, and that the fix is to compress both features and item-item structure toward what actually predicts user preference. It proposes IBMRec, built on the Information Bottleneck principle, with a feature-level module (FIB) that maximizes the mutual information between multimedia representations and ratings while minimizing dependence on the raw pre-trained features, and a graph-level module (GIB) that rebuilds the item-item graph from preference affinity and then minimizes dependence between the original and refined graphs. On the Amazon Clothing, Sports, and Baby datasets, IBMRec is reported to outperform all baselines, with relative NDCG@20 gains of 14.25%, 8.53%, and 8.96% over the strongest baseline, MICOR, and to improve several existing recommenders when attached to them. The intended upshot is that deliberately discarding information from rich multimedia inputs makes recommenders more accurate, and the Information Bottleneck supplies a task-agnostic way to decide what to discard.

What carries the argument

The load-bearing machinery is the Information Bottleneck objective restated for recommender data: maximize $I(R; Z)$ while minimizing $I(M; Z)$ for features, and maximize task-relevant structure while minimizing $I(S'; S)$ for the item-item graph. Direct computation is handled in three moves: first, $I(R; U, V, Z)$ is decomposed as $I(R; U, V) + I(R; Z | U, V)$ and each term is replaced by a tractable lower bound (BPR ranking loss and a conditional noise-contrastive loss); second, the intractable minimization of $I(M; Z)$ is approximated by HSIC computed in a reproducing kernel Hilbert space with RBF kernels; third, the graph-level objective relaxes $I(S'; S)$ to node-level $I(Y'; Y)$ between the embeddings produced by the original and refined graphs, and again uses HSIC. The preference-guided edge-masking MLP, which keeps each item-item edge with probability $w_{ij}$ learned from denoised multimedia and collaborative representations, is what injects collaborative information into the graph refinement before the HSIC penalty is applied.

What would settle it

On a small dataset, compute a tight variational or exact upper bound of $I(S'; S)$ alongside the HSIC($Y'$, $Y$) surrogate the paper minimizes across training; if the surrogate falls while the true bound stays flat or rises, the graph-level module is not bottlenecking graph information. A simpler experiment: hold the preference-guided edge-masking network fixed and delete only the HSIC regularization; unchanged accuracy would mean the gains come from the masking MLP rather than from Information Bottleneck compression.

Watch

Extended reading notes

Core claim

The central claim is that the Information Bottleneck principle ported to multimedia recommendation removes irrelevant multimedia features and thereby improves ranking accuracy. The paper treats pre-trained multimedia features $M$ as the input, the rating matrix $R$ as the task signal, and the learned multimedia representation $Z$ plus the refined item-item graph $S'$ as the bottleneck variables, optimizing $\max I(R; Z) - \alpha I(M; Z)$ at the feature level and $\max I(R; U, V, S') - \beta I(S'; S)$ at the graph level. Because direct mutual information estimation is intractable, the paper decomposes $I(R; U, V, Z)$ into lower bounds realized by a BPR ranking loss and a conditional noise-contrastive loss, and replaces the minimization terms with HSIC, a kernel-based statistical independence criterion. The experiments report consistent gains over all baselines on three datasets, and ablations show that removing either IB module degrades performance, while adding more modalities helps IBMRec even though it hurts its VLATTICE backbone. The discovery, on the paper's own terms, is that less information, chosen by an Information Bottleneck objective, can be more signal for the recommendation task.

Load-bearing premise

The load-bearing premise is that measuring the statistical dependence between the original and denoised item graphs can be swapped for measuring the dependence between the embeddings those graphs produce; if that swap does not upper-bound the actual graph-level information, the graph-level module is not enforcing an information bottleneck.

Editorial extensions

If this is right

  • Multimedia recommenders that rely on generic pre-trained extractors leave accuracy on the table: compressing the extracted features toward rating-predictive content should improve ranking without retraining the extractor.
  • Structure-level denoising matters even after feature-level compression, because an item-item graph built from noisy features amplifies the noise through graph convolution; GIB targets that amplification directly.
  • IBMRec behaves as a plugin: attaching FIB and GIB to backbones such as VBPR, VLightGCN, LATTICE, and VLATTICE is reported to improve each one, with the largest relative gains on the noisiest or weakest settings.
  • More modalities become useful once noise is removed: the paper reports that combining visual and textual features hurts its VLATTICE backbone but helps IBMRec, suggesting the bottleneck turns modality fusion from a liability into an asset.

Reading between the lines

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

  • An implication the paper leaves implicit is that the mechanism can be stress-tested with controlled noise: adding synthetic irrelevant dimensions to image or text features should make IBMRec's advantage over non-denoised baselines grow monotonically, and if it does not, the reported gains are not caused by removing irrelevant features.
  • The paper does not isolate the contribution of the preference-guided masking MLP from the HSIC penalty; an ablation that keeps the MLP but removes only the HSIC terms would tell whether Information Bottleneck compression or the added collaborative edge mask drives the results.
  • Because the GIB relaxation measures embedding-level dependence rather than graph-level mutual information, a stricter estimator or an exact small-scale computation could reveal whether the method truly minimizes graph information or merely regularizes embeddings.
  • The reported gains are concentrated on sparse, small datasets; testing on denser or larger-scale multimedia benchmarks would show whether the denoising benefit survives when collaborative signal alone is already strong.
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 / 5 minor

Summary. The paper proposes IBMRec, a multimedia recommendation model built around the Information Bottleneck (IB) principle. Two regularizers are added to a LATTICE-style backbone: a feature-level IB module (FIB) that maximizes mutual information between the learned multimedia representation and the rating signal while using HSIC to reduce dependence between the representation and the original pre-trained multimedia features, and a graph-level IB module (GIB) that first re-weights the item-item graph with a preference-guided MLP and then uses an HSIC term to reduce dependence between embeddings obtained from the original and refined graphs. Experiments on Amazon Clothing, Sports, and Baby report consistent accuracy gains over existing baselines, e.g., 14.25%, 8.53%, and 8.96% relative NDCG@20 improvements over MICOR, and improved performance when IBMRec is attached to several backbones.

Significance. The empirical portion is the strongest part of the manuscript: the reported gains are consistent across three datasets, several baselines, and multiple backbones, and the paper includes ablation studies and hyper-parameter sensitivity analyses. If the IB interpretation were fully justified, the two-level denoising framework would be a useful and reasonably novel contribution to multimedia recommendation. However, the central theoretical derivation is not currently established. The graph-level IB objective relies on an unjustified equality between graph-level mutual information and a sum of node-level mutual informations, and the HSIC surrogate is not a controlled approximation of that objective. In addition, the ablations do not isolate the contribution of the HSIC regularizer from the contribution of the preference-guided edge-masking MLP. The paper therefore reports a promising empirical recipe, but the load-bearing claim that it implements Information Bottleneck denoising at both feature and graph levels needs substantial revision.

major comments (4)
  1. [§4.2.1, Eq. (18)] The derivation of the InfoNCE-style objective contains an unstated expectation condition. In Eq. (18), the step from the sum over negative samples to the factor (N-1) requires E_j[p(r_ai | z_j, a, i) / p(r_ai | a, i)] = 1, but this is not implied by the preceding definitions and is not generally true; p(r_ai | a, i) is defined through f(a,i), not as the marginal of p(r_ai | z_j, a, i) integrated over the batch distribution of z_j. Relatedly, Eq. (19) replaces p(r_ai | a, i) with an average over batch items z_j, which is a heuristic rather than a consequence of the probabilistic model. The final objective Eq. (20) is a familiar InfoNCE lower bound, but the derivation as written does not prove it. Please either state the missing condition, provide a correct derivation, or explicitly present Eq. (20) as a standard contrastive lower bound rather than as a quantity derived from Eqs. (17)-(19).
  2. [§4.3.2, Eq. (26)] The equality Min I(S'; S) = Σ_i I(Y'; Y) is asserted without proof and is not an identity or a bound. Mutual information between two graphs does not decompose into a sum of mutual informations between node embeddings produced by a GNN. As a concrete failure mode, if S' = S then the left-hand side is H(S) (for discrete graphs), while the right-hand side is I(Y; Y) = H(Y); these are generally different. Conversely, if S' and S are independent but Y' and Y are both computed from the same user-item graph A and the same multimedia features M, the two embedding sets can be strongly dependent even when the graph structures are independent. Because Eq. (26) is the step that converts the graph-level IB objective into a computable node-level objective, the graph-level IB claim is not supported without a derivation or a bound relating these quantities.
  3. [§4.3.2, Eq. (28) and Algorithm 1] The replacement of the graph-level objective by HSIC(Y'; Y) is not justified as an approximation of I(S'; S). HSIC on paired node embeddings measures dependence between embedding sets, not between the graph structures S' and S, and no upper bound or controlled-error statement connects the two. The notation is also internally inconsistent: Eq. (28) writes LGIB = \hat{HSIC}(S'; S') ≈ \hat{HSIC}(Y'; Y), while Algorithm 1 line 8 says \hat{HSIC}(S', S) and Eq. (30) writes \hat{HSIC}(\tilde Y'; \tilde Y). The authors should either prove a relation between the HSIC term and the graph-level IB objective, or explicitly reframe GIB as a heuristic regularizer that is only motivated by, rather than derived from, the Information Bottleneck principle.
  4. [§5.3.2, Table 4] The ablation study does not isolate the contribution of the HSIC term in GIB from the contribution of the preference-guided edge-masking MLP. The GIB module in Eqs. (23)-(24) and Eq. (28) contains two distinct components: the construction of S' through the MLP-gated mask, and the HSIC regularization on Y' and Y. The 'IBMRec-w/o GIB' variant removes both components together, so the improvement of the full model over this variant could be entirely due to the edge-masking MLP rather than to the information-bottleneck regularizer. Please add an ablation that keeps the preference-guided masking but removes only the HSIC term, and ideally also a variant that applies the HSIC term to the original graph S instead of the masked graph S'. Without such an ablation, the empirical evidence does not establish that graph-level Information Bottleneck denoising, rather than the architectural masking component, is responsible for the reported gains.
minor comments (5)
  1. [Throughout] There are several notation and naming inconsistencies: Eq. (28) uses HSIC(S'; S') while Algorithm 1 uses HSIC(S', S) and Eq. (30) uses a garbled \hat{HSIC}(\tilde Y'; \tilde Y); the baseline name appears as both MICRO and MICOR; SGFP in the related work is SGFD in the experiments. Please unify these.
  2. [§4.1.2, Eq. (12)] The symbol Y is used both for the item embedding matrix after user-item graph convolution and for the final fused item representation in Eq. (12); this makes Eq. (27) and Eq. (30) harder to follow. Please use distinct symbols.
  3. [§5.4.2, Figs. 5 and 6] The text says 'As shown in Fig. 6' when discussing the FIB parameters (alpha, sigma^2), but the corresponding panel is Fig. 5; the GIB parameter discussion also refers to Fig. 6, which is likely the intended panel. Please correct the cross-references.
  4. [§5.1.2] The evaluation reports means over 10 repeated runs but no standard deviations or significance tests. Given that the reported relative improvements are substantial, adding variance or significance information would strengthen the comparison.
  5. [§1, contributions] The contribution statement claims to be the first to introduce the IB principle to multimedia recommendation. Given existing IB-based recommendation and graph IB works cited in the paper itself (e.g., Refs. [23], [46]), please soften or qualify the novelty claim.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation: FIB/GIB objectives are standard MI approximations; the unproven graph-level relaxation is a correctness gap, not a self-referential reduction.

full rationale

The paper's central claim—that IBMRec removes task-irrelevant multimedia features via the Information Bottleneck principle—is implemented as concrete losses: the I(R;Z) maximization is lower-bounded through BPR (Eq. 16) and InfoNCE (Eq. 20), and the I(Z;M) minimization is replaced by the HSIC regularizer (Eq. 22). These are known mutual-information surrogates from external literature; none of them is defined in terms of the reported NDCG or Recall, so the empirical gains are not forced by construction. The graph-level module is the weakest part: Eq. 26 asserts I(S';S)=sum_i I(Y';Y) without proof, and Eq. 28 substitutes HSIC(Y';Y) (with an apparent S';S' typo), so the graph-level IB claim is not established by the equations as written. This is a derivation-quality/validity issue, not circularity, because the masked graph and the HSIC regularizer are not the same object as the evaluation metric, and the Table 4 ablations compare variants rather than refitting the target. Self-citations [43], [44], [47] appear in related work and do not carry the central argument. Overall: no significant circularity; the score reflects the minor non-load-bearing self-citations and the unproven GIB relaxation noted under correctness risk rather than circular reasoning.

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

The central claim rests on a series of modeling choices: pretrained features contain separable irrelevant components; HSIC with an RBF kernel is a good proxy for mutual information; the InfoNCE lower bound is a valid objective for the multimedia preference score; and the graph-level IB can be replaced by node-level HSIC. The first three are defensible but unverified in this paper; the last is asserted without proof. The free parameters alpha, beta, and sigma squared are tuned per dataset on validation, and tau is not reported. No new physical or conceptual entities are introduced.

free parameters (6)
  • alpha (FIB loss coefficient) = 3.0 / 1.0 / 1.0 on Clothing / Sports / Baby
    Tuned on validation (Fig. 5); balances the recommendation loss and the HSIC feature bottleneck. Per-dataset optimal values vary.
  • beta (GIB loss coefficient) = 20 / 0.5 / 0.5 on Clothing / Sports / Baby
    Tuned on validation (Fig. 6); balances the recommendation loss and the HSIC graph bottleneck. The range spans two orders of magnitude across datasets.
  • sigma squared (RBF kernel width) = 0.15 to 0.30 (e.g., 0.25 / 0.20 / 0.15 for FIB)
    Controls the sharpness of the HSIC kernel; selected per dataset in the sensitivity analysis.
  • tau (InfoNCE temperature) = not reported
    Used in Eq. 20 for the density-ratio approximation; the paper gives no value or sensitivity analysis.
  • kNN Top-K for item-item graph = not reported
    Eq. 9 sparsifies each modality graph by keeping Top-K neighbors; the value is not stated and is likely inherited from LATTICE.
  • Number of graph convolution layers L = not reported
    Eqs. 6 and 11 stack L layers; L is not specified in the experimental settings.
assumptions (6)
  • standard math Mutual information chain rule I(R; U, V, Z) = I(R; U, V) + I(R; Z | U, V)
    Used in Eq. 15; standard identity, so it is not a burden.
  • domain assumption BPR pairwise ranking loss is a valid surrogate for maximizing I(R; U, V)
    The paper derives a lower bound in Eq. 16 and then substitutes the common BPR loss; the equivalence is not formal.
  • ad hoc to paper For negative samples z_j, E_j[p(r_ai | z_j, a, i)] = p(r_ai | a, i)
    Needed in Eq. 18 to convert the InfoNCE-style sum into log(1 + (N-1) p(r | z_i) / p(r)); not stated or tested.
  • ad hoc to paper Graph-level mutual information I(S'; S) equals sum_i I(Y'; Y)
    Eq. 26 asserts this equality without proof; it is the basis of GIB optimization.
  • domain assumption HSIC with RBF kernel approximates the mutual information terms in the IB objectives
    Used in Eqs. 22 and 28 following [26] and [39]; the fidelity of this proxy for high-dimensional recommendation embeddings is not evaluated.
  • domain assumption Pretrained multimedia features are a mixture of task-relevant and task-irrelevant components that can be separated by compression
    This is the paper's motivating premise (Section 1) and is necessary for FIB and GIB to improve rather than harm accuracy.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Less is More: Information Bottleneck Denoised Multimedia Recommendation." pith.science (2026). https://pith.science/paper/HJ7AUXVD

@misc{pith2026250112175,
  author       = {Pith},
  title        = {Pith review of: Less is More: Information Bottleneck Denoised Multimedia Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HJ7AUXVD}},
  note         = {Machine review of arXiv:2501.12175}
}
read the original abstract

Empowered by semantic-rich content information, multimedia recommendation has emerged as a potent personalized technique. Current endeavors center around harnessing multimedia content to refine item representation or uncovering latent item-item structures based on modality similarity. Despite the effectiveness, we posit that these methods are usually suboptimal due to the introduction of irrelevant multimedia features into recommendation tasks. This stems from the fact that generic multimedia feature extractors, while well-designed for domain-specific tasks, can inadvertently introduce task-irrelevant features, leading to potential misguidance of recommenders. In this work, we propose a denoised multimedia recommendation paradigm via the Information Bottleneck principle (IB). Specifically, we propose a novel Information Bottleneck denoised Multimedia Recommendation (IBMRec) model to tackle the irrelevant feature issue. IBMRec removes task-irrelevant features from both feature and item-item structure perspectives, which are implemented by two-level IB learning modules: feature-level (FIB) and graph-level (GIB). In particular, FIB focuses on learning the minimal yet sufficient multimedia features. This is achieved by maximizing the mutual information between multimedia representation and recommendation tasks, while concurrently minimizing it between multimedia representation and pre-trained multimedia features. Furthermore, GIB is designed to learn the robust item-item graph structure, it refines the item-item graph based on preference affinity, then minimizes the mutual information between the original graph and the refined one. Extensive experiments across three benchmarks validate the effectiveness of our proposed model, showcasing high performance, and applicability to various multimedia recommenders.

Figures

Figures reproduced from arXiv: 2501.12175 by the authors.

Figure 1
Figure 1. Illustration of irrelevant features on multimedia-based recommen [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Illustration of the optimization objectives. (a) Traditional multimedia recommendation methods only maximize the mutual information between [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Overall framework of our proposed IBMRec , we elaborate two Information Bottleneck learning modules to reduce the impact of irrelevant multimedia features for recommendations tasks: (a) Feature-level IB learning: aiming to remove task-irrelevant multimedia features in representation; (b) Graph-level IB learning: aiming to remove task-irrelevant graph structures to enhance recommendations. where DS is the degree matr… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Performance comparisons under different modalities. and VLATTICE (with the textual feature) across three datasets, the latter consistently exhibits better perfor￾mance. This suggests that the textual modality imparts additional information that enhances the overall qua…
Figure 5
Figure 5. Figure 5: Impact of different FIB loss parameters α and σ 2 . 10 15 20 25 0.15 0.20 0.25 0.30 2 0.0845 0.0868 0.0873 0.0869 0.0865 0.0877 0.0882 0.0875 0.0867 0.0881 0.0891 0.0890 0.0863 0.0876 0.0890 0.0886 0.085 0.086 0.087 0.088 0.089 (a) Clothing 0.25 0.5 1.0 2.0 0.15 0.20 0…
Figure 6
Figure 6. Figure 6: Impact of different GIB loss parameters β and σ 2 . [5] J. Chen, H. Zhang, X. He, L. Nie, W. Liu, and T.-S. Chua. Attentive collaborative filtering: Multimedia recommendation with item￾and component-level attention. In SIGIR, pages 335–344, 2017. [6] P. Cheng, W. Hao, …

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Hierarchical Graph Information Bottleneck for Multi-Behavior Recommendation

    cs.IR 2025-07 conditional novelty 6.0 of 10

    HGIB improves multi-behavior recommendation by adding information-bottleneck preservation and compression losses plus a learnable graph refinement encoder to hierarchical models, with reported gains on academic and in...

Reference graph

Works this paper leans on

52 extracted references · 48 canonical work pages · cited by 1 Pith paper

  1. [23]

    D. Liu, P . Cheng, H. Zhu, Z. Dong, X. He, W. Pan, and Z. Ming. Mitigating confounding bias in recommendation via information bottleneck. In RecSys, pages 351–360, 2021

    work page 2021

  2. [1]

    A. A. Alemi, I. Fischer, J. V . Dillon, and K. Murphy. Deep variational information bottleneck. In ICLR, 2017

    work page 2017

  3. [2]

    S. Bang, P . Xie, H. Lee, W. Wu, and E. Xing. Explaining a black-box by using a deep variational information bottleneck approach. In AAAI, volume 35, pages 11396–11404, 2021

    work page 2021

  4. [3]

    M. I. Belghazi, A. Baratin, S. Rajeshwar, S. Ozair, Y. Bengio, A. Courville, and D. Hjelm. Mutual information neural estimation. In ICML, pages 531–540. PMLR, 2018

    work page 2018

  5. [4]

    C. Chen, M. Zhang, Y. Liu, and S. Ma. Neural attentional rating regression with review-level explanations. In WWW, pages 1583– 1592, 2018. JOURNAL OF XXX 11 1.0 2.0 3.0 4.0 0.15 0.20 0.25 0.30 2 0.0863 0.0868 0.0869 0.0863 0.0872 0.0871 0.0881 0.0866 0.0873 0.0872 0.0891 0.0864 0.0871 0.0880 0.0885 0.0855 0.0855 0.0860 0.0865 0.0870 0.0875 0.0880 0.0885 0...

    work page 2018

  6. [5]

    J. Chen, H. Zhang, X. He, L. Nie, W. Liu, and T.-S. Chua. Attentive collaborative filtering: Multimedia recommendation with item- and component-level attention. In SIGIR, pages 335–344, 2017

    work page 2017

  7. [6]

    Cheng, W

    P . Cheng, W. Hao, S. Dai, J. Liu, Z. Gan, and L. Carin. Club: A contrastive log-ratio upper bound of mutual information. In ICML, pages 1779–1788. PMLR, 2020

    work page 2020

  8. [7]

    Deldjoo, M

    Y. Deldjoo, M. Schedl, P . Cremonesi, and G. Pasi. Recommender systems leveraging multimedia content. ACM Computing Surveys (CSUR), 53(5):1–38, 2020

    work page 2020

Show all 52 references
  1. [8]

    Devlin, M.-W

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova. Bert: Pre- training of deep bidirectional transformers for language under- standing. In NAACL-HLT, pages 4171–4186, 2019

    2019

  2. [9]

    S. Ding, W. Du, L. Ding, J. Zhang, L. Guo, and B. An. Robust multi-agent communication with graph information bottleneck optimization. IEEE TP AMI, 2023

    2023

  3. [10]

    X. Du, Z. Wu, F. Feng, X. He, and J. Tang. Invariant representation learning for multimedia recommendation. In MM, pages 619–628, 2022

    2022

  4. [11]

    Espinosa Zarlenga, P

    M. Espinosa Zarlenga, P . Barbiero, G. Ciravegna, G. Marra, F. Giannini, M. Diligenti, Z. Shams, F. Precioso, S. Melacci, A. Weller, et al. Concept embedding models: Beyond the accuracy- explainability trade-off. NeurIPS, 35:21400–21413, 2022

    2022

  5. [12]

    Gitiaux and H

    X. Gitiaux and H. Rangwala. Fair representations by compression. In AAAI, volume 35, pages 11506–11515, 2021

    2021

  6. [13]

    Goldfeld, E

    Z. Goldfeld, E. v. d. Berg, K. Greenewald, I. Melnyk, N. Nguyen, B. Kingsbury, and Y. Polyanskiy. Estimating information flow in deep neural networks. In ICML, pages 2299–2308, 2018

    2018

  7. [14]

    Gretton, O

    A. Gretton, O. Bousquet, A. Smola, and B. Sch ¨olkopf. Measuring statistical dependence with hilbert-schmidt norms. In International conference on algorithmic learning theory , pages 63–77. Springer, 2005

    2005

  8. [15]

    Gronowski, W

    A. Gronowski, W. Paul, F. Alajaji, B. Gharesifard, and P . Burlina. Classification utility, fairness, and compactness via tunable in- formation bottleneck and r ´enyi measures. IEEE Transactions on Information Forensics and Security, 2023

    2023

  9. [16]

    He and J

    R. He and J. McAuley. Vbpr: visual bayesian personalized ranking from implicit feedback. In AAAI, pages 144–150, 2016

    2016

  10. [17]

    X. He, K. Deng, X. Wang, Y. Li, Y. Zhang, and M. Wang. Lightgcn: Simplifying and powering graph convolution network for recom- mendation. In SIGIR, pages 639–648, 2020

    2020

  11. [18]

    X. He, L. Liao, H. Zhang, L. Nie, X. Hu, and T.-S. Chua. Neural collaborative filtering. In WWW, pages 173–182, 2017

    2017

  12. [19]

    S. Hu, Z. Lou, X. Yan, and Y. Ye. A survey on information bottleneck. IEEE TP AMI, 2024

    2024

  13. [20]

    Huang, H

    S. Huang, H. Li, Q. Li, C. Zheng, and L. Liu. Pareto invariant representation learning for multimedia recommendation. arXiv preprint arXiv:2308.04706, 2023

    2023 arXiv

  14. [21]

    T. N. Kipf and M. Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907 , 2016

    2016 arXiv

  15. [22]

    X. Lei, X. Qian, and G. Zhao. Rating prediction based on social sentiment from textual reviews. IEEE TMM, 18(9):1910–1921, 2016

    1910

  16. [24]

    F. Liu, H. Chen, Z. Cheng, L. Nie, and M. Kankanhalli. Semantic- guided feature distillation for multimodal recommendation. In MM, pages 6567–6575, 2023

    2023

  17. [25]

    Q. Liu, S. Wu, and L. Wang. Deepstyle: Learning user preferences for visual recommendation. In SIGIR, pages 841–844, 2017

    2017

  18. [26]

    W.-D. K. Ma, J. Lewis, and W. B. Kleijn. The hsic bottleneck: Deep learning without back-propagation. In AAAI, volume 34, pages 5085–5092, 2020

    2020

  19. [27]

    McAuley, C

    J. McAuley, C. Targett, Q. Shi, and A. Van Den Hengel. Image- based recommendations on styles and substitutes. In SIGIR, pages 43–52, 2015

    2015

  20. [28]

    Y.-I. Moon, B. Rajagopalan, and U. Lall. Estimation of mutual information using kernel density estimators. Physical Review E , 52(3):2318, 1995

    1995

  21. [29]

    A. v. d. Oord, Y. Li, and O. Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018

    2018 arXiv

  22. [30]

    Reimers and I

    N. Reimers and I. Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. In EMNLP, pages 3980–3990, 2019

    2019

  23. [31]

    Rendle, C

    S. Rendle, C. Freudenthaler, Z. Gantner, and L. Schmidt-Thieme. Bpr: Bayesian personalized ranking from implicit feedback. In UAI, pages 452–461, 2009

    2009

  24. [32]

    A. M. Saxe, Y. Bansal, J. Dapello, M. Advani, A. Kolchinsky, B. D. Tracey, and D. D. Cox. On the information bottleneck JOURNAL OF XXX 12 theory of deep learning. Journal of Statistical Mechanics: Theory and Experiment, 2019(12):124020, 2019

    2019

  25. [33]

    Shuai, K

    J. Shuai, K. Zhang, L. Wu, P . Sun, R. Hong, M. Wang, and Y. Li. A review-aware graph contrastive learning framework for recommendation. In SIGIR, pages 1283–1293, 2022

    2022

  26. [34]

    Simonyan and A

    K. Simonyan and A. Zisserman. Very deep convolutional net- works for large-scale image recognition. 2015

    2015

  27. [35]

    Tishby, F

    N. Tishby, F. C. Pereira, and W. Bialek. The information bottleneck method. arXiv preprint physics/0004057, 2000

    2000 arXiv

  28. [36]

    Tishby and N

    N. Tishby and N. Zaslavsky. Deep learning and the information bottleneck principle. In 2015 ieee information theory workshop (itw) , pages 1–5. IEEE, 2015

    2015

  29. [37]

    J.-P . Vert, K. Tsuda, and B. Sch¨olkopf. A primer on kernel methods. Kernel methods in computational biology, 47:35–70, 2004

    2004

  30. [38]

    J. Wang, M. Luo, J. Li, Y. Lin, Y. Dong, J. S. Dong, and Q. Zheng. Empower post-hoc graph explanations with information bottle- neck: A pre-training and fine-tuning perspective. In KDD, pages 2349–2360, 2023

    2023

  31. [39]

    Z. Wang, T. Jian, A. Masoomi, S. Ioannidis, and J. Dy. Revisiting hilbert-schmidt information bottleneck for adversarial robustness. NeurIPS, 34:586–597, 2021

    2021

  32. [40]

    Y. Wei, X. Wang, L. Nie, X. He, and T.-S. Chua. Graph-refined con- volutional network for multimedia recommendation with implicit feedback. In MM, pages 3541–3549, 2020

    2020

  33. [41]

    Y. Wei, X. Wang, L. Nie, X. He, R. Hong, and T.-S. Chua. Mmgcn: Multi-modal graph convolution network for personalized recom- mendation of micro-video. In MM, pages 1437–1445, 2019

    2019

  34. [42]

    J. Wu, X. Wang, F. Feng, X. He, L. Chen, J. Lian, and X. Xie. Self- supervised graph learning for recommendation. In SIGIR, pages 726–735, 2021

    2021

  35. [43]

    L. Wu, Y. Yang, L. Chen, D. Lian, R. Hong, and M. Wang. Learning to transfer graph embeddings for inductive graph based recommendation. In SIGIR, pages 1211–1220, 2020

    2020

  36. [44]

    L. Wu, Y. Yang, K. Zhang, R. Hong, Y. Fu, and M. Wang. Joint item recommendation and attribute inference: An adaptive graph convolutional network approach. In SIGIR, pages 679–688, 2020

    2020

  37. [45]

    S. Wu, F. Sun, W. Zhang, X. Xie, and B. Cui. Graph neural networks in recommender systems: a survey. ACM Computing Surveys, 55(5):1–37, 2022

    2022

  38. [46]

    T. Wu, H. Ren, P . Li, and J. Leskovec. Graph information bottle- neck. NeurIPs, 33:20437–20448, 2020

    2020

  39. [47]

    Y. Yang, L. Wu, K. Zhang, R. Hong, H. Zhou, Z. Zhang, J. Zhou, and M. Wang. Hyperbolic graph learning for social recommenda- tion. IEEE TKDE, 2023

    2023

  40. [48]

    Z. Yi, X. Wang, I. Ounis, and C. Macdonald. Multi-modal graph contrastive learning for micro-video recommendation

  41. [49]

    J. Yu, T. Xu, Y. Rong, Y. Bian, J. Huang, and R. He. Recognizing predictive substructures with subgraph information bottleneck. IEEE TP AMI, 46(3):1650–1663, 2021

    2021

  42. [50]

    Zhang, X

    C. Zhang, X. Zhou, Y. Wan, X. Zheng, K.-W. Chang, and C.-J. Hsieh. Improving the adversarial robustness of nlp models by information bottleneck. arXiv preprint arXiv:2206.05511, 2022

    2022 arXiv

  43. [51]

    Zhang, Y

    J. Zhang, Y. Zhu, Q. Liu, S. Wu, S. Wang, and L. Wang. Mining latent structures for multimedia recommendation. In MM, pages 3872–3880, 2021

    2021

  44. [52]

    Zhang, Y

    J. Zhang, Y. Zhu, Q. Liu, M. Zhang, S. Wu, and L. Wang. Latent structure mining with contrastive modality fusion for multimedia recommendation. IEEE TKDE, 2022. Y onghui Yangis currently pursuing a Ph.D. de- gree at Hefei University of Technology, China. He obtained his master...

    2022

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.