Pith. sign in

REVIEW 3 major objections 6 minor 55 references

Hierarchical Graph Information Bottleneck for Multi-Behavior Recommendation

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read An information-bottleneck objective with learnable graph refinement improves target-behavior prediction in multi-behavior recommenders, yielding consistent gains over thirteen baselines and a positive 15-day online A/B test.

desk verdict Solid empirical wrapper with a shaky IB theory; the L2 substitution in Eq (5)-(7) is unproved, but the method works and deserves review. read the letter →

arxiv 2507.15395 v1 pith:VADX43WA submitted 2025-07-21 cs.IR

Hengyu Zhang , Chunxu Shen , Xiangguo Sun , Jie Tan , Yanchao Tan , Yu Rong , Hong Cheng , Lingling Yi This is my paper · ORCID
classification cs.IR
keywords multi-behaviorrecommendationinformationbottleneckgraphneuralnetworksRefinementEncoderHSICcompressionInfoNCEpreservationnegativetransferhierarchical
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 sets out to show that the two persistent failure modes of multi-behavior recommenders—imbalanced interaction distributions and negative transfer from noisy auxiliary behaviors—can be handled by one information-theoretic objective imposed on the whole hierarchical encoder stack. The proposed HGIB layer adds a preservation loss (InfoNCE between intermediate and final embeddings), a compression loss (HSIC between consecutive encoder levels), and a learnable Graph Refinement Encoder that prunes low-confidence user-item edges before aggregation. If the claim holds, practitioners gain a model-agnostic recipe for exploiting abundant view and cart signals without letting them corrupt purchase prediction. The support is consistent: relative gains of 6–15 percent in HR@10 and NDCG@10 over the strongest baseline on three public datasets, double-digit gains on two industrial-scale datasets, and statistically significant CTR and sharing-rate lifts in a 15-day online A/B test.

What carries the argument

The load-bearing object is the HGIB objective of Eq. (15): $\mathcal{L}_{rec}$ (cross-entropy on target behavior), $\mathcal{L}_{pres}$ (InfoNCE between the final fused output $O$ and each encoder output), $\mathcal{L}_{comp}$ (a sum of HSIC terms over consecutive encoder levels), and $\mathcal{L}_{reg}$ (L2 regularization). It is mounted on a hierarchical backbone—unified graph encoder, behavior-specific encoders, and behavior-component encoders over intersection and difference graphs, fused by target attention—so every level of the hierarchy receives the same preserve-versus-compress instruction. The named mechanism, Graph Refinement Encoder (GRE), computes edge weights $w_{uv} = \sigma(e_u^T e_v)$, zeroes weights at or below threshold $\tau$, applies Gumbel-softmax reparameterization, and then aggregates with LightGCN. The machinery's job is to keep target-relevant information flowing down the hierarchy while discarding auxiliary-behavior noise and pruning the noisy edges themselves before they ever reach the predictor.

What would settle it

On the Taobao dataset, record the empirical differential entropy of the final embedding matrix and the squared L2 norm of all trainable parameters at every checkpoint during HGIB training. The derivation requires that shrinking the parameter norm pushes $-H(E_L)$ upward (equivalently pushes $H(E_L)$ downward); if the two quantities move apart across checkpoints or random seeds, the 'can be treated as' substitution in Section 4.2 fails and the implemented loss is not the information-bottleneck bound the paper claims.

Watch

Extended reading notes

Core claim

The central claim is that in a hierarchical multi-behavior recommender, the right representation for the target behavior is the one that keeps every encoder level predictive of that behavior while compressing whatever the previous level supplied, and that this objective can be made tractable. The paper derives a per-layer information-bottleneck objective and argues that maximizing its lower bound reduces to minimizing cross-entropy on the final fused representation, InfoNCE between the final and intermediate embeddings, and an L2 parameter penalty, with the compression side approximated by HSIC between neighboring layers. On top of this, the Graph Refinement Encoder scores each interaction edge as $\sigma(e_u^T e_v)$, drops edges below a threshold $\tau$, and uses Gumbel-softmax reparameterization so the discrete pruning stays end-to-end trainable. Instantiated on a parallel-paradigm backbone with unified, behavior-specific, and behavior-component encoders fused by target attention, HGIB outperforms thirteen baselines on Taobao, Tmall, and Jdata, improves both AutoDCS and MULE when integrated into them, and shows gains on two billion-interaction industrial datasets plus a live 15-day A/B test.

Load-bearing premise

The load-bearing premise is the Section 4.2 step that replaces the negative entropy term $-H(E_L)$ with an L2-norm penalty on trainable parameters: the paper says it 'can be treated as' such a penalty without proof or a cited bound, and the claim that the trained loss maximizes the IB objective depends entirely on that substitution.

Editorial extensions

If this is right

  • The same four-term objective can be attached to any hierarchical multi-behavior model; the compatibility experiments show gains on both the cascading-style AutoDCS and the parallel MULE, so the benefit does not depend on the paper's own backbone.
  • Auxiliary behaviors become safer to use: because GRE prunes noisy edges and $\mathcal{L}_{comp}$ suppresses irrelevant transfer, systems can retain abundant view and cart interactions without paying the negative-transfer penalty they normally incur.
  • The hierarchy becomes inspectable: information abundance decreases with encoder depth, so shallow layers store broad interaction structure and deeper layers keep only target-relevant signal, giving a training diagnostic for whether a backbone is learning the intended funnel.
  • Compression strength matters: the coefficient $\beta$ has a sweet spot near 50, meaning practitioners must tune how much auxiliary information is discarded rather than treating more compression as always better.
  • If the online result generalizes, the framework gives production recommenders a direct lever on business-defined target behaviors: sharing rate rose 1.65 percent while click-through rose 0.32 percent, so the gains are not just engagement-wide.

Reading between the lines

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

  • Beyond recommendation, the same hierarchical preserve-and-compress scheme is a generic recipe for any multi-view graph encoder, for example heterogeneous social or knowledge graphs; the paper tests only user-item behaviors, so the breadth of the mechanism is an extrapolation.
  • The derivation's entropy-to-L2 step, flagged in Section 4.2 as 'can be treated as,' is the natural place to stress-test the theory: an ablation that estimates the entropy term directly rather than replacing it with weight decay would show whether the implemented loss is truly the IB bound or just a useful set of regularizers.
  • Because the industrial comparison pits HGIB against MULE alone, a reader can infer that the production gain depends on how much denoising the incumbent ranker already performs; on a baseline with aggressive behavior-specific dropout, the reported 0.32 percent CTR and 1.65 percent SSR lift could shrink, and that is testable.
  • The information-abundance plots suggest a cheap screening test for adoption: if a team's current backbone already shows high information abundance at shallow layers, the bottleneck losses may add little, whereas collapsed shallow embeddings indicate the framework is likely to help.
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. The paper proposes HGIB, a model-agnostic framework for multi-behavior recommendation that combines a hierarchical graph encoder with an information-bottleneck-style objective. The objective is implemented as a target-behavior cross-entropy loss, an InfoNCE-based preservation loss, an HSIC-based compression loss, L2 regularization, and a Graph Refinement Encoder that prunes edges via a learnable threshold. The authors evaluate HGIB on Taobao, Tmall, and Jdata, on two industrial datasets, and in a 15-day online A/B test, reporting consistent improvements over state-of-the-art baselines.

Significance. If the empirical results are robust, HGIB is practically valuable: it reports large relative gains across three public datasets, holds up in a large-scale online A/B test, and is compatible with existing hierarchical backbones. The authors also release source code, which supports reproducibility. The theoretical claim that the loss in Eq. (15) is derived from the information bottleneck objective is not currently established because of the unproved replacement of an entropy term by L2 regularization; the empirical recipe may still be useful, but the paper should either supply the missing bound or reframe the contribution as a heuristic objective guided by IB ideas.

major comments (3)
  1. [Section 4.2, Eq. (5)-(7)] The derivation replaces -H(E_L) with L2-norm regularization on trainable parameters without any proof or inequality. In Eq. (5), the lower bound on I(E_l;Y) ends with the term -H(E_L); Eq. (7) then substitutes lambda*||Theta||^2 for this term. No bound of the form lambda*||Theta||^2 >= H(E_L) or lambda*||Theta||^2 <= H(E_L) is given, and the paper provides no argument that one follows from the architecture. Consequently, Eq. (7) is not shown to be equivalent to maximizing the lower bound in Eq. (5), and the implemented objective in Eq. (15) is not shown to optimize the IB objective in Eq. (4). The naming and theoretical motivation of HGIB therefore rest on an unproved step. Please supply a valid bound or explicitly state that Eq. (15) is an IB-inspired heuristic.
  2. [Section 5.2, Table 2] The central empirical claim relies on Table 2, but no standard deviations, confidence intervals, or number of runs are reported. The asterisks indicating p<0.05 are not accompanied by a description of the statistical test (e.g., paired test, number of seeds, user-level or sample-level comparison). The same issue affects Tables 3, 4, and 6. Without this information, the relative improvements, especially the smaller Jdata gains (6.12% HR, 4.12% NDCG over AutoDCS), cannot be assessed for significance.
  3. [Section 5.4, Table 4] The model-agnostic claim is weakened by the AutoDCS compatibility results on Jdata, where the relative improvements are 0.39% in HR@10 and 0.26% in NDCG@10, which are far smaller than the other compatibility gains. The text states that 'HGIB consistently achieves significant performance gains across different backbone models' without qualification. Please either provide statistical evidence that these small gains are significant or temper the generality claim.
minor comments (6)
  1. [Section 4.4, Eq. (10)] The Gumbel-softmax relaxation of the hard threshold is not described in detail; as written, the hard threshold function is not differentiable and it is unclear how the reparameterization is applied to the edge weights.
  2. [Section 5.5] The sensitivity analysis varies alpha and beta but not the GRE threshold tau (fixed at 0.05), even though the GRE module is a stated contribution; please report the effect of tau.
  3. [References] References [41] and [42] are the same paper (Graph Information Bottleneck for Subgraph Recognition); one should be removed or replaced.
  4. [Table 1] Tmall and Taobao are listed with the same number of items (11,953); please verify this statistic.
  5. [Section 5.6.1] The industrial experiments use leave-one-out evaluation while the public experiments use full ranking; the difference should be stated explicitly in the main text to avoid implying that Table 2 and Table 6 are directly comparable.
  6. [Section 4.6.1] The claimed O(B^3 L) HSIC complexity with B=1024 may still be substantial; please report actual training time or memory cost to support the claim that overhead is not significant.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the Eq. (5)-(7) entropy-to-L2 substitution is unproved but is a theoretical-validity gap, not a by-construction equivalence.

full rationale

The paper's central derivation chain is Eq. (4)-(16). The lower-bound algebra in Eq. (5) is a valid sequence of inequalities; replacing -H(Y|E_L) by cross-entropy and I(E_L;E_l) by InfoNCE are standard lower bounds. The only questionable step is Section 4.2's assertion that the third term -H(E_L) "can be treated as the L2-norm regularization of the trainable parameters." No inequality relating entropy to parameter norm is given, so the implemented objective in Eq. (7)/(15) is not actually proven to maximize the claimed IB lower bound. That is an unsupported approximation or theoretical-validity gap, but it is not circular: HGIB's loss is not defined in terms of its reported metrics, nor are the reported gains fitted from the loss. The empirical comparisons in Tables 2, 4, and 6 and the online A/B test are measured against independent baselines, so the performance claims do not reduce to the objective's construction. Self-citations (refs 22, 23, 44, 45, 46) appear only as baselines or as examples of existing attention/GCN techniques and do not carry the derivation. The RQ6 information-abundance analysis is post-hoc explanation using HGIB's own embeddings; it does not constitute a predicted quantity and does not feed back into the loss. Hence no step reduces by construction to its inputs.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The method leans on several approximations and ad-hoc substitutions. The central empirical claims do not depend on the theoretical derivation, but the paper's stated contribution as an IB framework rests on assumptions that are not rigorously justified. No new physical or theoretical entities are introduced.

free parameters (4)
  • alpha (preservation coefficient) = 1.0 (Taobao, Tmall), 0.5 (Jdata)
    Controls weight of InfoNCE preservation loss; chosen per dataset via experiments in Section 5.5.
  • beta (compression coefficient) = 50
    Controls weight of HSIC compression loss; default set for all datasets, selected in Section 5.5.
  • tau (edge pruning threshold) = 0.05
    Threshold below which edges are pruned in GRE; set as default for all datasets without per-dataset tuning.
  • lambda/gamma (regularization coefficient) = 0.1
    L2 regularization coefficient in the final loss.
assumptions (5)
  • standard math InfoNCE lower bound: I(E_L; E_l) >= log(N) - L_CL(E_L,E_l)
    From van den Oord et al. [28]; used in Eq (6) to convert mutual information to contrastive loss.
  • domain assumption HSIC approximates mutual information and can stand in for I(E_l; E_{l-1}) minimization
    Standard approximation in IB literature, cited from [21,29,40]; used in Eq (9).
  • domain assumption Any hierarchical multi-behavior model can be abstracted as a sequence of encoders E_{l+1} = Encoder_{l+1}(E_l)
    Assumed in Section 4.1.1 to make the IB formulation general; questionable for cascading models with different graph inputs.
  • ad hoc to paper The negative entropy term -H(E_L) can be replaced by L2-norm regularization on parameters
    Stated in Section 4.2 without proof; not a valid bound, making the derivation heuristic.
  • domain assumption Edge weights sigma(e_u^T e_v) above threshold tau indicate genuine interactions
    Underlies GRE in Eq (10); assumes dot-product similarity of embeddings correlates with interaction reliability.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Hierarchical Graph Information Bottleneck for Multi-Behavior Recommendation." pith.science (2026). https://pith.science/paper/VADX43WA

@misc{pith2026250715395,
  author       = {Pith},
  title        = {Pith review of: Hierarchical Graph Information Bottleneck for Multi-Behavior Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VADX43WA}},
  note         = {Machine review of arXiv:2507.15395}
}
abstract

In real-world recommendation scenarios, users typically engage with platforms through multiple types of behavioral interactions. Multi-behavior recommendation algorithms aim to leverage various auxiliary user behaviors to enhance prediction for target behaviors of primary interest (e.g., buy), thereby overcoming performance limitations caused by data sparsity in target behavior records. Current state-of-the-art approaches typically employ hierarchical design following either cascading (e.g., view$\rightarrow$cart$\rightarrow$buy) or parallel (unified$\rightarrow$behavior$\rightarrow$specific components) paradigms, to capture behavioral relationships. However, these methods still face two critical challenges: (1) severe distribution disparities across behaviors, and (2) negative transfer effects caused by noise in auxiliary behaviors. In this paper, we propose a novel model-agnostic Hierarchical Graph Information Bottleneck (HGIB) framework for multi-behavior recommendation to effectively address these challenges. Following information bottleneck principles, our framework optimizes the learning of compact yet sufficient representations that preserve essential information for target behavior prediction while eliminating task-irrelevant redundancies. To further mitigate interaction noise, we introduce a Graph Refinement Encoder (GRE) that dynamically prunes redundant edges through learnable edge dropout mechanisms. We conduct comprehensive experiments on three real-world public datasets, which demonstrate the superior effectiveness of our framework. Beyond these widely used datasets in the academic community, we further expand our evaluation on several real industrial scenarios and conduct an online A/B testing, showing again a significant improvement in multi-behavior recommendations. The source code of our proposed HGIB is available at https://github.com/zhy99426/HGIB.

Figures

Figures reproduced from arXiv: 2507.15395 by the authors.

Figure 1
Figure 1. Illustration of the hierarchical model design in [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 3
Figure 3. Illustration of the proposed HGIB, with three types of behavior as examples (view as [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Performance with the coefficient 𝛼 for L𝑝𝑟𝑒𝑠 . 1 10 50100 1 10 50100 The coefficient β 0.212 0.215 0.217 0.220 0.223 0.225 HR@10 0.118 0.119 0.120 0.121 0.122 0.123 NDCG@10 (a) Taobao 1 10 50100 1 10 50100 The coefficient β 0.228 0.232 0.236 0.240 0.244 0.248 0.252 HR@10 0.124 0.126 0.128 0.129 0.131 0.132 NDCG@10 (b) Tmall 1 10 50100 1 10 50100 The coefficient β 0.630 0.636 0.642 0.648 0.654 0.660 HR@10 0.450 0.455… view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: Performance with the coefficient 𝛽 for L𝑐𝑜𝑚𝑝 . The comparative analysis of hyperparameters is illustrated in [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Comparison of information abundance of the em [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

55 extracted references · 48 canonical work pages

  1. [1]

    Chong Chen, Weizhi Ma, Min Zhang, Zhaowei Wang, Xiuqiang He, Chenyang Wang, Yiqun Liu, and Shaoping Ma. 2021. Graph Heterogeneous Multi-Relational Recommendation. In AAAI. AAAI Press, 3958–3966

    work page 2021

  2. [2]

    Chong Chen, Min Zhang, Yongfeng Zhang, Weizhi Ma, Yiqun Liu, and Shaoping Ma. 2020. Efficient Heterogeneous Collaborative Filtering without Negative Sampling for Recommendation. In AAAI. AAAI Press, 19–26

    work page 2020

  3. [3]

    Zhiyong Cheng, Sai Han, Fan Liu, Lei Zhu, Zan Gao, and Yuxin Peng. 2023. Multi-Behavior Recommendation with Cascading Graph Convolution Networks. In WWW. ACM, 1181–1189. RecSys ’25, September 22–26, 2025, Prague, Czech Republic Hengyu Zhang et al

    work page 2023

  4. [4]

    Jingtao Ding, Guanghui Yu, Xiangnan He, Yuhan Quan, Yong Li, Tat-Seng Chua, Depeng Jin, and Jiajie Yu. 2018. Improving Implicit Recommender Systems with View Data. In IJCAI. ijcai.org, 3343–3349

    work page 2018

  5. [5]

    Shuheng Fang, Kangfei Zhao, Yu Rong, Jeffrey Xu Yu, and Zhixun Li. 2025. All- in-One: Heterogeneous Interaction Modeling for Cold-Start Rating Prediction . In ICDE. IEEE Computer Society, Los Alamitos, CA, USA, 1537–1550. doi:10. 1109/ICDE65448.2025.00119

    arXiv 2025

  6. [6]

    Chen Gao, Xiangnan He, Dahua Gan, Xiangning Chen, Fuli Feng, Yong Li, Tat- Seng Chua, and Depeng Jin. 2019. Neural Multi-task Recommendation from Multi-behavior Data. In ICDE. IEEE, 1554–1557

    work page 2019

  7. [7]

    Smola, and Bernhard Schölkopf

    Arthur Gretton, Olivier Bousquet, Alexander J. Smola, and Bernhard Schölkopf

  8. [8]

    Long Guo, Lifeng Hua, Rongfei Jia, Binqiang Zhao, Xiaobo Wang, and Bin Cui. 2019. Buying or Browsing?: Predicting Real-time Purchasing Intent using Attention-based Deep Network with Multiple Behavior. In KDD. ACM, 1984– 1992

    work page 2019

Show all 55 references
  1. [9]

    Wei Guo, Chang Meng, Enming Yuan, Zhicheng He, Huifeng Guo, Yingxue Zhang, Bo Chen, Yaochen Hu, Ruiming Tang, Xiu Li, and Rui Zhang. 2023. Compressed Interaction Graph based Framework for Multi-behavior Recommendation. In WWW. ACM, 960–970

    2023

  2. [10]

    Xingzhuo Guo, Junwei Pan, Ximei Wang, Baixu Chen, Jie Jiang, and Mingsheng Long. 2024. On the Embedding Collapse when Scaling up Recommendation Models. In ICML. OpenReview.net

    2024

  3. [11]

    Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yong-Dong Zhang, and Meng Wang. 2020. LightGCN: Simplifying and Powering Graph Convolution Network for Recommendation. In SIGIR. ACM, 639–648

    2020

  4. [12]

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

    2017

  5. [13]

    Eric Jang, Shixiang Gu, and Ben Poole. 2017. Categorical Reparameterization with Gumbel-Softmax. In ICLR (Poster). OpenReview.net

    2017

  6. [14]

    Bowen Jin, Chen Gao, Xiangnan He, Depeng Jin, and Yong Li. 2020. Multi- behavior Recommendation with Graph Convolutional Networks. In SIGIR. ACM, 659–668

    2020

  7. [15]

    Artus Krohn-Grimberghe, Lucas Drumond, Christoph Freudenthaler, and Lars Schmidt-Thieme. 2012. Multi-relational matrix factorization using bayesian personalized ranking for social network data. In WSDM. ACM, 173–182

    2012

  8. [16]

    Seunghan Lee, Geonwoo Ko, Hyun-Je Song, and Jinhong Jung. 2024. MuLe: Multi-Grained Graph Learning for Multi-Behavior Recommendation. In CIKM. ACM, 1163–1173

    2024

  9. [17]

    Dugang Liu, Pengxiang Cheng, Hong Zhu, Zhenhua Dong, Xiuqiang He, Weike Pan, and Zhong Ming. 2021. Mitigating Confounding Bias in Recommendation via Information Bottleneck. In RecSys. ACM, 351–360

    2021

  10. [18]

    Dugang Liu, Pengxiang Cheng, Hong Zhu, Zhenhua Dong, Xiuqiang He, Weike Pan, and Zhong Ming. 2023. Debiased Representation Learning in Recommenda- tion via Information Bottleneck. Trans. Recomm. Syst. 1, 1 (2023), 1–27

    2023

  11. [19]

    Dugang Liu, Shenxian Xian, Yuhao Wu, Chaohua Yang, Xing Tang, Xiuqiang He, and Zhong Ming. 2024. AutoDCS: Automated Decision Chain Selection in Deep Recommender Systems. In SIGIR. ACM, 956–965

    2024

  12. [20]

    Larson, and Alan Hanjalic

    Babak Loni, Roberto Pagano, Martha A. Larson, and Alan Hanjalic. 2016. Bayesian Personalized Ranking with Multi-Channel User Feedback. In RecSys. ACM, 361– 364

    2016

  13. [21]

    Kurt Wan-Duo Ma, J. P. Lewis, and W. Bastiaan Kleijn. 2020. The HSIC Bottleneck: Deep Learning without Back-Propagation. In AAAI. AAAI Press, 5085–5092

    2020

  14. [22]

    Chang Meng, Chenhao Zhai, Yu Yang, Hengyu Zhang, and Xiu Li. 2023. Parallel Knowledge Enhancement based Framework for Multi-behavior Recommendation. In CIKM. ACM, 1797–1806

    2023

  15. [23]

    Chang Meng, Hengyu Zhang, Wei Guo, Huifeng Guo, Haotian Liu, Yingxue Zhang, Hongkun Zheng, Ruiming Tang, Xiu Li, and Rui Zhang. 2023. Hierarchical Projection Enhanced Multi-behavior Recommendation. In KDD. ACM, 4649– 4660

    2023

  16. [24]

    Qi Pi, Guorui Zhou, Yujing Zhang, Zhe Wang, Lejian Ren, Ying Fan, Xiaoqiang Zhu, and Kun Gai. 2020. Search-based User Interest Modeling with Lifelong Sequential Behavior Data for Click-Through Rate Prediction. In CIKM. ACM, 2685–2692

    2020

  17. [25]

    Steffen Rendle, Christoph Freudenthaler, Zeno Gantner, and Lars Schmidt-Thieme

  18. [26]

    Yu Rong, Xiao Wen, and Hong Cheng. 2014. A Monte Carlo algorithm for cold start recommendation. In WWW. ACM, 327–336

    2014

  19. [27]

    Kipf, Peter Bloem, Rianne van den Berg, Ivan Titov, and Max Welling

    Michael Sejr Schlichtkrull, Thomas N. Kipf, Peter Bloem, Rianne van den Berg, Ivan Titov, and Max Welling. 2018. Modeling Relational Data with Graph Con- volutional Networks. In ESWC (Lecture Notes in Computer Science, Vol. 10843) . Springer, 593–607

    2018

  20. [28]

    Aäron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. Representation Learning with Contrastive Predictive Coding. CoRR abs/1807.03748 (2018)

    2018 arXiv

  21. [29]

    Zifeng Wang, Tong Jian, Aria Masoomi, Stratis Ioannidis, and Jennifer G. Dy. 2021. Revisiting Hilbert-Schmidt Information Bottleneck for Adversarial Robustness. In NeurIPS. 586–597

    2021

  22. [30]

    Chunyu Wei, Jian Liang, Di Liu, and Fei Wang. 2022. Contrastive Graph Structure Learning via Information Bottleneck for Recommendation. In NeurIPS

    2022

  23. [31]

    Wei Wei, Chao Huang, Lianghao Xia, Yong Xu, Jiashu Zhao, and Dawei Yin. 2022. Contrastive Meta Learning with Behavior Multiplicity for Recommendation. In WSDM. ACM, 1120–1128

    2022

  24. [32]

    Lianghao Xia, Chao Huang, Yong Xu, Peng Dai, Mengyin Lu, and Liefeng Bo. 2021. Multi-Behavior Enhanced Recommendation with Cross-Interaction Collaborative Relation Modeling. In ICDE. IEEE, 1931–1936

    2021

  25. [33]

    Lianghao Xia, Chao Huang, Yong Xu, Peng Dai, Bo Zhang, and Liefeng Bo

  26. [34]

    Lianghao Xia, Yong Xu, Chao Huang, Peng Dai, and Liefeng Bo. 2021. Graph Meta Network for Multi-Behavior Recommendation. In SIGIR. ACM, 757–766

    2021

  27. [35]

    Junsong Xie, Yonghui Yang, Zihan Wang, and Le Wu. 2024. Learning Fair Repre- sentations for Recommendation via Information Bottleneck Principle. In IJCAI. ijcai.org, 2469–2477

    2024

  28. [36]

    Jingcao Xu, Chaokun Wang, Cheng Wu, Yang Song, Kai Zheng, Xiaowei Wang, Changping Wang, Guorui Zhou, and Kun Gai. 2023. Multi-behavior Self- supervised Learning for Recommendation. In SIGIR. ACM, 496–505

    2023

  29. [37]

    Mingshi Yan, Zhiyong Cheng, Chen Gao, Jing Sun, Fan Liu, Fuming Sun, and Haojie Li. 2024. Cascading Residual Graph Convolutional Network for Multi- Behavior Recommendation. ACM Trans. Inf. Syst. 42, 1 (2024), 10:1–10:26

    2024

  30. [38]

    Mingshi Yan, Zhiyong Cheng, Jing Sun, Fuming Sun, and Yuxin Peng. 2023. MB-HGCN: A Hierarchical Graph Convolutional Network for Multi-behavior Recommendation. CoRR abs/2306.10679 (2023)

    2023 arXiv

  31. [39]

    Yonghui Yang, Le Wu, Zhuangzhuang He, Zhengwei Wu, Richang Hong, and Meng Wang. 2025. Less is More: Information Bottleneck Denoised Multimedia Recommendation. CoRR abs/2501.12175 (2025)

    2025 arXiv

  32. [40]

    Yonghui Yang, Le Wu, Zihan Wang, Zhuangzhuang He, Richang Hong, and Meng Wang. 2024. Graph Bottlenecked Social Recommendation. In KDD. ACM, 3853–3862

    2024

  33. [41]

    Junchi Yu, Tingyang Xu, Yu Rong, Yatao Bian, Junzhou Huang, and Ran He

  34. [42]

    Junchi Yu, Tingyang Xu, Yu Rong, Yatao Bian, Junzhou Huang, and Ran He. 2021. Graph Information Bottleneck for Subgraph Recognition. In ICLR

    2021

  35. [43]

    Chenhao Zhai, Chang Meng, Yu Yang, Kexin Zhang, Xuhao Zhao, and Xiu Li. 2025. Combinatorial Optimization Perspective based Framework for Multi-behavior Recommendation. In KDD (1). ACM, 1891–1902

    2025

  36. [44]

    Hengyu Zhang, Junwei Pan, Dapeng Liu, Jie Jiang, and Xiu Li. 2024. Deep Pattern Network for Click-Through Rate Prediction. In SIGIR. ACM, 1189–1199

    2024

  37. [45]

    Hengyu Zhang, Chunxu Shen, Xiangguo Sun, Jie Tan, Yu Rong, Chengzhi Piao, Hong Cheng, and Lingling Yi. 2025. Adaptive Graph Integration for Cross- Domain Recommendation via Heterogeneous Graph Coordinators. In SIGIR. ACM, 1860–1869

    2025

  38. [46]

    Hengyu Zhang, Enming Yuan, Wei Guo, Zhicheng He, Jiarui Qin, Huifeng Guo, Bo Chen, Xiu Li, and Ruiming Tang. 2022. Disentangling Past-Future Modeling in Sequential Recommendation via Dual Networks. In CIKM. ACM, 2549–2558

    2022

  39. [47]

    Weifeng Zhang, Jingwen Mao, Yi Cao, and Congfu Xu. 2020. Multiplex Graph Neural Networks for Multi-behavior Recommendation. In CIKM. ACM, 2313– 2316

    2020

  40. [48]

    Zhe Zhao, Zhiyuan Cheng, Lichan Hong, and Ed Huai-hsin Chi. 2015. Improving User Topic Interest Profiles by Behavior Factorization. In WWW. ACM, 1406– 1416

    2015

  41. [49]

    Guorui Zhou, Na Mou, Ying Fan, Qi Pi, Weijie Bian, Chang Zhou, Xiaoqiang Zhu, and Kun Gai. 2019. Deep Interest Evolution Network for Click-Through Rate Prediction. In AAAI. AAAI Press, 5941–5948

    2019

  42. [50]

    Guorui Zhou, Xiaoqiang Zhu, Chengru Song, Ying Fan, Han Zhu, Xiao Ma, Yanghui Yan, Junqi Jin, Han Li, and Kun Gai. 2018. Deep Interest Network for Click-Through Rate Prediction. In KDD. ACM, 1059–1068

    2018

  43. [51]

    Xi Zhu, Fake Lin, Ziwei Zhao, Tong Xu, Xiangyu Zhao, Zikai Yin, Xueying Li, and Enhong Chen. 2025. Multi-Behavior Recommendation with Personalized Directed Acyclic Behavior Graphs. ACM Trans. Inf. Syst. 43, 1 (2025), 20:1–20:30

    2025

  44. [2005]

    In ALT (Lecture Notes in Computer Science, Vol

    Measuring Statistical Dependence with Hilbert-Schmidt Norms. In ALT (Lecture Notes in Computer Science, Vol. 3734) . Springer, 63–77

  45. [2009]

    BPR: Bayesian Personalized Ranking from Implicit Feedback. In UAI. AUAI Press, 452–461

  46. [2020]

    In SIGIR

    Multiplex Behavioral Relation Learning for Recommendation via Memory Augmented Transformer Network. In SIGIR. ACM, 2397–2406

  47. [2021]

    Graph Information Bottleneck for Subgraph Recognition. In ICLR. OpenRe- view.net

Pith tools

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