Pith. sign in

REVIEW 4 major objections 6 minor 2 cited by

Breaking Information Cocoons: A Hyperbolic Framework for Balancing Exploration and Exploitation in Recommender Systems

T0 review · 4 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read HERec claims that aligning LLM text profiles with collaborative signals in hyperbolic space lets a single recommender system raise both relevance and diversity, reducing information cocoons.

desk verdict Solid hyperbolic recommender with a real alignment contribution, but the Dasgupta-cost hierarchy claim is unsupported and the headline 'consistently outperforms' is too strong. read the letter →

arxiv 2411.13865 v4 pith:LTOOSXZX submitted 2024-11-21 cs.IR cs.AIcs.CLcs.LG

Qiyao Ma , Menglin Yang , Mingxuan Ju , Tong Zhao , Neil Shah , Rex Ying This is my paper · ORCID
classification cs.IRcs.AIcs.CLcs.LG
keywords informationcocoonsexploration-exploitationtrade-offhyperbolicrecommendersystemsalignmenthierarchicalclusteringDasgupta'scostLLMsemanticprofilesdiversitymetrics
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 tries to break information cocoons by making recommender systems explore as well as exploit. Its proposal, HERec, fuses semantic profiles generated by LLMs with collaborative user-item signals directly in hyperbolic space, and then builds a hierarchy tree over the learned embeddings that users can adjust with two controls. The authors argue that hyperbolic geometry gives the alignment an adaptive gradient behavior that preserves the latent preference hierarchy, and that the tree permits a principled exploration-exploitation balance. If the claims hold, a single model can improve recommendation accuracy and diversity at the same time, which existing Euclidean and hyperbolic baselines do not achieve.

What carries the argument

The load-bearing machinery is twofold. First, the hyperbolic alignment loss $\ell_{align}(i) = d_H^2(h_i, s_i)$ pulls collaborative hyperbolic embeddings $h_i$ toward LLM-derived semantic embeddings $s_i$ projected into the same Lorentz model; Proposition 1 argues that the gradient magnitude in hyperbolic space is inversely proportional to node norm, so fine-grained (large-norm) nodes receive smaller updates and abstract (small-norm) nodes receive larger updates, preserving hierarchy. Second, the hierarchical representation structure builds a binary tree bottom-up via hyperbolic k-means clustering (Algorithm 1), creating pseudo-cluster nodes as centroids at each layer; the paper motivates this with Dasgupta's cost, which favors binary trees that cut edges low, and then exposes two user controls: temperature $\tau$ (the fraction of recommendations replaced) and layer $l$ (where in the tree to sample replacement items from ancestor clusters).

What would settle it

Run HERec's exploration mechanism with the hierarchy tree replaced by either a Euclidean k-means tree or a random binary tree over the same embeddings and compare diversity and utility; if the gains persist, the hyperbolic hierarchy is not the active ingredient. Also compute Dasgupta's cost on the produced tree: the paper claims cost-optimal hierarchy discovery, yet Algorithm 1 never evaluates this cost, so demonstrating that a cheaper tree exists would show the stated mechanism is not what the algorithm actually optimizes.

Watch

Extended reading notes

Core claim

The paper proposes HERec, a recommender that does hyperbolic graph collaborative filtering on user-item interactions, generates semantic user and item profiles with LLMs, and aligns those profiles to the hyperbolic embeddings with a distance-based alignment loss. The central claim is that this joint hyperbolic alignment, guided by an adaptive gradient property (Proposition 1: gradient magnitude scales as $\|\hat{x} - \hat{y}\|/(\|x\|(1-\cos\theta))$), preserves hierarchical preference structure, and that a hierarchy tree built by hyperbolic k-means over the final embeddings supports a user-adjustable exploration-exploitation trade-off. Empirically the paper reports consistent wins over Euclidean and hyperbolic baselines on Amazon-books, Yelp, and Google-reviews, with up to 5.49% improvement in utility (Recall and NDCG) and 11.39% improvement in diversity (distance diversity, Shannon entropy, and expected popularity complement), including larger gains on tail items. The paper claims this is the first model to excel at both utility and diversity simultaneously.

Load-bearing premise

The headline balancing result depends on the assumption that the binary tree produced by hyperbolic k-means over the final user and item embeddings corresponds to the users' true latent preference hierarchy, so that replacing a recommendation with an item sampled from an ancestor cluster is meaningful exploration rather than random substitution.

Editorial extensions

If this is right

  • If HERec is right, recommender systems can raise diversity without a separate post-hoc reranking stage that usually sacrifices accuracy.
  • The semantic alignment specifically helps tail and cold-start items, because text profiles fill in when interaction history is sparse.
  • Users or platforms can tune the exploration-exploitation balance by setting $\tau$ and layer $l$, with small-layer replacement favoring diversity and large-layer replacement favoring utility.
  • The hyperbolic gradient property implies that alignment updates respect the hierarchy: broad preferences are adjusted globally while niche preferences are adjusted locally.

Reading between the lines

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

  • The Dasgupta-cost motivation is not actually implemented in Algorithm 1: the algorithm is fixed-$k$ hyperbolic k-means with $k=2$, and no Dasgupta cost is computed, so the 'hyperparameter-free' and 'optimizing Dasgupta's cost' claims are weaker than stated; an actual greedy cost-minimizing split would be a testable improvement.
  • The two user controls could be personalized: a learned policy could set $\tau$ and layer per user from engagement signals, something the paper leaves implicit.
  • The hyperbolic alignment idea is transferable to any domain with text plus an interaction graph, such as news, job matching, or scientific papers, where the same alignment loss could be applied even without the hierarchy tree.
  • The hierarchy tree itself could enable continuous exploration by interpolating along tree paths rather than discrete branch replacement, which the paper does not explore.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes HERec, a hyperbolic graph-LLM recommender system that combines hyperbolic graph collaborative filtering with LLM-derived semantic profiles, aligned in Lorentz space via a distance loss. It also proposes a hierarchical clustering step that builds a binary tree over the learned embeddings, intended to let users balance exploration and exploitation by replacing part of the recommendation list with items sampled from ancestor clusters at a chosen layer. The paper claims consistent outperformance over Euclidean and hyperbolic baselines on both utility and diversity metrics, with up to 5.49% utility and 11.39% diversity improvement, and claims that the hierarchy is discovered by optimizing Dasgupta's cost without predefined hyperparameters.

Significance. If the claims held, this would be a notable result: a single model improving both utility and diversity while offering user-controllable exploration would be practically useful and would extend hyperbolic collaborative filtering with semantic grounding. The paper ships an open-source implementation, reports standard deviations over 10 runs (Table 5), and includes a head/tail analysis (Table 2) that is a genuine strength. However, the two headline contributions are weakened by internal evidence: Table 1 contradicts the 'consistently outperforms' statement on specific diversity metrics, and the hierarchy-tree mechanism is not shown to optimize Dasgupta's cost or to deliver the claimed exploration-exploitation balance beyond a single anecdotal case study and one τ=0.5 sweep. The empirical core (hyperbolic alignment + margin loss) appears defensible, but the paper's broad claims need substantial revision.

major comments (4)
  1. [§5.2, Table 1] The statement that HERec 'consistently outperforms' all baselines in both utility and diversity is contradicted by the paper's own Table 1. On Google-reviews, HICF achieves Div@10 = 0.3262 while HERec achieves 0.3185 (HERec is second-best); on Yelp, HGCF achieves EPC@10 = 0.8718 while HERec achieves 0.8716; and on Google-reviews, SimGCL achieves NDCG@10 = 0.0784 while HERec achieves 0.0772. These are not isolated rounding effects given the reported standard deviations in Table 5. The abstract and §5.2 should be rephrased to claim 'state-of-the-art on most metrics' or to provide a significance test that justifies a broader claim.
  2. [§4.5, Appendix C.1] The hierarchy tree is described as being 'theoretically optimized by Dasgupta's cost' and as 'hyperparameter-free', but Algorithm 1 in Appendix C.1 is ordinary hyperbolic k-means: it takes embeddings X and a fixed proportion k=2, randomly initializes centroids, assigns points by nearest hyperbolic distance, and replaces points by centroids. No similarity graph, edge weights, or Dasgupta-cost objective appear anywhere in the algorithm or its convergence criterion. Furthermore, k=2 and the maximum number of layers L=log_k(|X|) are fixed inputs, so the 'without predefined hyperparameters' claim in §4.5 is not supported. The paper needs either a genuine derivation connecting Algorithm 1 to Dasgupta's cost, or a removal of that claim and an alternative characterization of the tree quality.
  3. [§5.4, Appendix C.3] The exploration-exploitation mechanism is validated only by a single τ=0.5 sweep in Figure 4 (no error bars, no comparison against random replacement or established diversity-enhancing methods such as MMR or DPP-based reranking) and one anecdotal case study in Appendix C.3. The claim that this mechanism 'effectively mitigates information cocoons' is therefore not substantiated. The authors should provide a more controlled evaluation of the tree-based sampling, e.g., compare it against random exploration at equivalent τ values and report standard deviations across runs.
  4. [§5.2 vs. §5.4] The paper never states whether Table 1's diversity results are obtained with the exploration mechanism active (i.e., τ>0) or with pure exploitation (τ=0). If Table 1 uses τ=0, then the diversity gains come from the learned embeddings, not from the hierarchy tree, and the second contribution's empirical support is limited to Figure 4. If Table 1 uses τ>0, then τ is a tuned hyperparameter and the 'hyperparameter-free' description of the hierarchy mechanism is misleading. The default experimental setting must be clarified, and results for both settings should be reported.
minor comments (6)
  1. [Table 1] The header contains a typo: 'Diveristy' should be 'Diversity'.
  2. [§4.1] The phrase 'hyperbolic messaging passing' should be 'hyperbolic message passing'.
  3. [Appendix D.2] Although the section is titled 'Statistical Significance Testing', it only reports standard deviations; no significance tests (e.g., paired t-tests or bootstrap confidence intervals) are performed. The title and discussion should be adjusted to reflect what is actually reported.
  4. [§4.5] The statement that 'the optimal tree is required to be binary' conflates Dasgupta's cost with a property of binary trees; Dasgupta's cost is defined for hierarchical clusterings and does not by itself imply that the optimal tree must be binary in the sense used here. Rephrase to avoid the unsupported implication.
  5. [Appendix C.2] The complexity analysis is not fully clear: 'each k-means iteration takes O(i)' is vague, and the O(N^2/2) term for pairwise distance computation should be stated as O(N^2) (constant factor omitted). The overall complexity bound should be derived more carefully.
  6. [Conclusion] There is a typo: 'a5.49%' should be 'a 5.49%'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: HERec's headline results are established against external baselines and held-out data, and no fitted constant or self-citation chain is renamed as a prediction.

full rationale

Walking the claimed derivation chain, I find no step in which a 'prediction' is equivalent, by the paper's own equations or definitions, to a fitted input or to a self-citation. The hyperbolic alignment loss (Eq. 13) is a training objective, and Proposition 1 is an analytic gradient-magnitude computation with stated approximations and an error bound (Appendix A); neither is fitted to the reported gains. The main utility/diversity claims (Tables 1-2) are benchmarked against external Euclidean and hyperbolic baselines on public datasets, with standard deviations reported, so the headline results are not forced by construction. The hierarchical layer experiment (Sec. 5.4, Fig. 4) shows a trade-off that the paper explicitly describes as aligning with the model's design; although the direction of the trade-off is structurally related to how layer l controls ancestor-cluster sampling (Sec. 4.5), this is presented as a design verification rather than an independent empirical derivation. The profile-decoder test (Appendix E) reconstructs the same profiles that generated the semantic embeddings used in alignment, but it is a relative comparison across embedding sources and is not used as the central evidence of recommendation quality. The paper reuses prior components from the authors' own work (LLM profile extraction [20,23]; HICF margin/negative sampling [36]), but these are standard building blocks and are not the justification for the claimed superiority. The discrepancy between the 'optimizing Dasgupta's cost' language and Algorithm 1's hyperbolic k-means is a correctness/validity concern, not a circularity: the algorithm is not identical to the claimed objective, so no result is reduced to an input by construction.

Assumptions & free parameters 4 free parameters · 5 assumptions · 2 invented entities

The framework depends on a small set of fitted or hand-chosen parameters (alignment weight, k, temperature, layer), a domain assumption that hyperbolic geometry matches user-item hierarchies, and an unproven connection between Dasgupta's cost and the implemented k-means algorithm. The invented entities are internal clustering constructs with no independent falsifiable handles.

free parameters (4)
  • layer proportion k = 2
    Fixed to 2 because the paper asserts Dasgupta's cost implies binary optimal trees, but Algorithm 1 does not optimize Dasgupta's cost; k is a hand-chosen clustering hyperparameter despite the hyperparameter-free claim.
  • temperature tau = user-adjustable; 0.5 used in Section 5.4
    Controls the fraction of top recommendations replaced with items from other hierarchy branches. It is a user input, not learned, and is central to the exploration-exploitation trade-off.
  • hierarchy layer l = user-adjustable; 5 used in the example
    Selects the ancestor cluster used for exploration. The experimental sweep in Figure 4 varies this parameter, so the claimed trade-off is parameter-dependent.
  • alignment weight = grid searched in {1e-3, 1e-2, 1e-1}
    Weights the hyperbolic alignment loss in Section 4.3; chosen by grid search in Section B.1, so the reported performance depends on this tuning.
assumptions (5)
  • domain assumption User-item networks follow a power-law or long-tail distribution that hyperbolic embeddings can capture.
    Motivates the whole hyperbolic setup; supported only by dataset tail statistics in Appendix B, not by a fitted model.
  • standard math Dasgupta's cost optimal trees can be assumed binary, so k=2 is a valid default.
    The paper cites Dasgupta for binary optimality, but the cited result does not imply that hyperbolic k-means with k=2 optimizes the cost; the connection is asserted, not derived.
  • domain assumption Embedding norms are sufficiently large for x0 approximately equal to ||x|| and y0 approximately equal to ||y|| in the gradient proof.
    Needed for Proposition 1 and Eq. 14; the paper argues 50-dimensional embeddings are large-norm, but the error-bound appendix is internally inconsistent.
  • ad hoc to paper Semantic embeddings from Euclidean text encoders can be MLP-projected into hyperbolic space and aligned with collaborative embeddings by distance loss without losing useful semantics.
    Underpins the semantic enhancement; no analysis of projection distortion or theoretical guarantee is provided.
  • domain assumption The hierarchy tree built from final embeddings reflects real preference groupings.
    The exploration mechanism samples from ancestor clusters; only a single case study in Appendix C.3 validates that these clusters correspond to meaningful preferences.
invented entities (2)
  • Pseudo-cluster nodes
    purpose: Centroids representing aggregated preference clusters at each hierarchy level; used to retrieve ancestor clusters for exploration.
    Internal algorithmic constructs from hyperbolic k-means; no external labels or independent validation beyond one case study.
  • Hierarchy tree
    purpose: Organizes users and items by preference granularity so users can switch between exploitation at deep layers and exploration at shallow layers.
    The tree is not compared against any ground-truth taxonomy or preference labels; its utility is assessed only through the system's own recommendation metrics.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Breaking Information Cocoons: A Hyperbolic Framework for Balancing Exploration and Exploitation in Recommender Systems." pith.science (2026). https://pith.science/paper/LTOOSXZX

@misc{pith2026241113865,
  author       = {Pith},
  title        = {Pith review of: Breaking Information Cocoons: A Hyperbolic Framework for Balancing Exploration and Exploitation in Recommender Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LTOOSXZX}},
  note         = {Machine review of arXiv:2411.13865}
}
read the original abstract

Modern recommender systems often create information cocoons, restricting users' exposure to diverse content. The central challenge is to balance content exploration and exploitation while allowing users to adjust their recommendation preferences. Ideally, this balance can be captured with a hierarchical representation, where depth search facilitates exploitation and breadth search enables exploration. However, existing approaches face two fundamental limitations: Euclidean methods struggle to capture hierarchical structures, while hyperbolic methods, despite their superior hierarchical modeling, lack semantic understanding of user and item profiles and fail to provide a principled mechanism for balancing exploration and exploitation. To address these challenges, we propose HERec, a hyperbolic framework that effectively balances exploration and exploitation in recommender systems. Our framework introduces two key innovations: (1) a semantic-enhanced hierarchical mechanism that aligns rich textual descriptions with collaborative information directly in hyperbolic space. Theoretical gradient analysis demonstrates that this alignment effectively leverages the underlying hyperbolic manifold structure, resulting in more accurate modeling of users and items; (2) an automatic hierarchical clustering mechanism by optimizing Dasgupta's cost, which discovers hierarchical structures without requiring predefined hyperparameters, enabling user-adjustable exploration-exploitation trade-offs. Extensive experiments demonstrate that HERec consistently outperforms both Euclidean and hyperbolic baselines, achieving up to 5.49% improvement in utility metrics and 11.39% increase in diversity metrics, effectively mitigating information cocoons.

Figures

Figures reproduced from arXiv: 2411.13865 by the authors.

Figure 1
Figure 1. The overall architecture of HERec. (i) Hyperbolic Graph Collaborative Filtering: Encodes [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Ablation study on model variants [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 4
Figure 4. Performance analysis across hierarchical structure layers. [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: A depiction of model prompt instruction. [PITH_FULL_IMAGE:figures/full_fig_p018_5.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

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

  1. Large Language Model Enhanced Recommender Systems: A Survey

    cs.IR 2024-12 unverdicted novelty 4.0 of 10

    A survey organizing LLM-enhanced recommender systems into knowledge, interaction, and model enhancement, and tracing a shift from explicit text to implicit embeddings and fine-tuned open-source LLMs.

  2. Hyperbolic Deep Learning for Foundation Models: A Survey

    cs.LG 2025-07 conditional novelty 1.0 of 10

    A structured survey of hyperbolic-geometry methods for foundation models, concluding the approach is promising but showing limited independent evidence at scale.

Reference graph

Works this paper leans on

49 extracted references · 28 canonical work pages · cited by 2 Pith papers

  1. [1]

    Tallrec: An effective and efficient tuning framework to align large language model with recommendation

    Keqin Bao, Jizhi Zhang, Yang Zhang, Wenjie Wang, Fuli Feng, and Xiangnan He. Tallrec: An effective and efficient tuning framework to align large language model with recommendation. In Proceedings of the 17th ACM Conference on Recommender Systems, pages 1007–1014, 2023

    2023

  2. [2]

    Riemannian adaptive optimization methods

    Gary Bécigneul and Octavian-Eugen Ganea. Riemannian adaptive optimization methods. arXiv preprint arXiv:1810.00760, 2018

    arXiv 2018

  3. [3]

    Lightgcl: Simple yet effective graph contrastive learning for recommendation

    Xuheng Cai, Chao Huang, Lianghao Xia, and Xubin Ren. Lightgcl: Simple yet effective graph contrastive learning for recommendation. arXiv preprint arXiv:2302.08191, 2023

    arXiv 2023

  4. [4]

    Novelty and diversity in recommender systems

    Pablo Castells, Neil Hurley, and Saul Vargas. Novelty and diversity in recommender systems. In Recommender systems handbook, pages 603–646. Springer, 2021

    work page 2021

  5. [5]

    Hyperbolic graph convolutional neural networks

    Ines Chami, Zhitao Ying, Christopher Ré, and Jure Leskovec. Hyperbolic graph convolutional neural networks. Advances in neural information processing systems, 32, 2019

    work page 2019

  6. [6]

    Neural collaborative reasoning

    Hanxiong Chen, Shaoyun Shi, Yunqi Li, and Yongfeng Zhang. Neural collaborative reasoning. In Proceedings of the Web Conference 2021, pages 1516–1527, 2021

    work page 2021

  7. [7]

    A cost function for similarity-based hierarchical clustering

    Sanjoy Dasgupta. A cost function for similarity-based hierarchical clustering. In Proceedings of the forty-eighth annual ACM symposium on Theory of Computing, pages 118–127, 2016

    work page 2016

  8. [8]

    Bert: Pre-training of deep bidirectional transformers for language understanding

    Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018

    arXiv 2018

Show all 49 references
  1. [9]

    Hakg: Hierarchy-aware knowledge gated network for recommendation

    Yuntao Du, Xinjun Zhu, Lu Chen, Baihua Zheng, and Yunjun Gao. Hakg: Hierarchy-aware knowledge gated network for recommendation. In Proceedings of the 45th international ACM SIGIR conference on Research and development in Information Retrieval, pages 1390–1400, 2022

    2022

  2. [10]

    Hyperbolic neural networks.Advances in neural information processing systems, 31, 2018

    Octavian Ganea, Gary Bécigneul, and Thomas Hofmann. Hyperbolic neural networks.Advances in neural information processing systems, 31, 2018

    2018

  3. [11]

    Lightgcn: Simplifying and powering graph convolution network for recommendation

    Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yongdong Zhang, and Meng Wang. Lightgcn: Simplifying and powering graph convolution network for recommendation. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval, pages...

    2020

  4. [12]

    How does message passing improve collaborative filtering? arXiv preprint arXiv:2404.08660, 2024

    Mingxuan Ju, William Shiao, Zhichun Guo, Yanfang Ye, Yozen Liu, Neil Shah, and Tong Zhao. How does message passing improve collaborative filtering? arXiv preprint arXiv:2404.08660, 2024

    2024 arXiv

  5. [13]

    Adam: A method for stochastic optimization

    Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014

    2014 arXiv

  6. [14]

    Matrix factorization techniques for recom- mender systems

    Yehuda Koren, Robert Bell, and Chris V olinsky. Matrix factorization techniques for recom- mender systems. Computer, 42(8):30–37, 2009

    2009

  7. [15]

    Hyperbolic geometry of complex networks

    Dmitri Krioukov, Fragkiskos Papadopoulos, Maksim Kitsak, Amin Vahdat, and Marián Boguná. Hyperbolic geometry of complex networks. Physical Review E—Statistical, Nonlinear, and Soft Matter Physics, 82(3):036106, 2010

    2010

  8. [16]

    Diversity in recommender systems–a survey

    Matevž Kunaver and Tomaž Požrl. Diversity in recommender systems–a survey. Knowledge- based systems, 123:154–162, 2017

    2017

  9. [17]

    Text is all you need: Learning language representations for sequential recommendation

    Jiacheng Li, Ming Wang, Jin Li, Jinmiao Fu, Xin Shen, Jingbo Shang, and Julian McAuley. Text is all you need: Learning language representations for sequential recommendation. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 1258–1267, 2023

    2023

  10. [18]

    Hyperbolic graph neural networks

    Qi Liu, Maximilian Nickel, and Douwe Kiela. Hyperbolic graph neural networks. Advances in neural information processing systems, 32, 2019. 10

    2019

  11. [19]

    Llm-rec: Personalized recommendation via prompting large language models

    Hanjia Lyu, Song Jiang, Hanqing Zeng, Yinglong Xia, Qifan Wang, Si Zhang, Ren Chen, Christopher Leung, Jiajie Tang, and Jiebo Luo. Llm-rec: Personalized recommendation via prompting large language models. arXiv preprint arXiv:2307.15780, 2023

    2023 arXiv

  12. [20]

    Xrec: Large language models for explainable recommendation

    Qiyao Ma, Xubin Ren, and Chao Huang. Xrec: Large language models for explainable recommendation. arXiv preprint arXiv:2406.02377, 2024

    2024 arXiv

  13. [21]

    Collaborative per- sonalized web recommender system using entropy based similarity measure

    Harita Mehta, Shveta Kundra Bhatia, Punam Bedi, and Veer Sain Dixit. Collaborative per- sonalized web recommender system using entropy based similarity measure. arXiv preprint arXiv:1201.4210, 2012

    2012 arXiv

  14. [22]

    Text and code embeddings by contrastive pre-training

    Arvind Neelakantan, Tao Xu, Raul Puri, Alec Radford, Jesse Michael Han, Jerry Tworek, Qiming Yuan, Nikolas Tezak, Jong Wook Kim, Chris Hallacy, et al. Text and code embeddings by contrastive pre-training. arXiv preprint arXiv:2201.10005, 2022

    2022 arXiv

  15. [23]

    Representation learning with large language models for recommendation

    Xubin Ren, Wei Wei, Lianghao Xia, Lixin Su, Suqi Cheng, Junfeng Wang, Dawei Yin, and Chao Huang. Representation learning with large language models for recommendation. In Proceedings of the ACM on Web Conference 2024, pages 3464–3475, 2024

    2024

  16. [24]

    Bpr: Bayesian personalized ranking from implicit feedback

    Steffen Rendle, Christoph Freudenthaler, Zeno Gantner, and Lars Schmidt-Thieme. Bpr: Bayesian personalized ranking from implicit feedback. arXiv preprint arXiv:1205.2618, 2012

    2012 arXiv

  17. [25]

    Hgcf: Hy- perbolic graph convolution networks for collaborative filtering

    Jianing Sun, Zhaoyue Cheng, Saba Zuberi, Felipe Pérez, and Maksims V olkovs. Hgcf: Hy- perbolic graph convolution networks for collaborative filtering. In Proceedings of the Web Conference 2021, pages 593–601, 2021

    2021

  18. [26]

    Knowledge based hyperbolic propagation

    Chang-You Tai, Chien-Kun Huang, Liang-Ying Huang, and Lun-Wei Ku. Knowledge based hyperbolic propagation. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1945–1949, 2021

    1945

  19. [27]

    Hierarchical latent context representation for context- aware recommendations

    Moshe Unger and Alexander Tuzhilin. Hierarchical latent context representation for context- aware recommendations. IEEE Transactions on Knowledge and Data Engineering , 34(7): 3322–3334, 2020

    2020

  20. [28]

    Rank and relevance in novelty and diversity metrics for recommender systems

    Saúl Vargas and Pablo Castells. Rank and relevance in novelty and diversity metrics for recommender systems. In Proceedings of the fifth ACM conference on Recommender systems, pages 109–116, 2011

    2011

  21. [29]

    Hyperml: A boosting metric learning approach in hyperbolic space for recommender systems

    Lucas Vinh Tran, Yi Tay, Shuai Zhang, Gao Cong, and Xiaoli Li. Hyperml: A boosting metric learning approach in hyperbolic space for recommender systems. In Proceedings of the 13th international conference on web search and data mining, pages 609–617, 2020

    2020

  22. [30]

    Neural graph collabo- rative filtering

    Xiang Wang, Xiangnan He, Meng Wang, Fuli Feng, and Tat-Seng Chua. Neural graph collabo- rative filtering. In Proceedings of the 42nd international ACM SIGIR conference on Research and development in Information Retrieval, pages 165–174, 2019

    2019

  23. [31]

    Llmrec: Large language models with graph augmentation for recommendation

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

    2024

  24. [32]

    Balancing exploration and exploitation with information and randomization

    Robert C Wilson, Elizabeth Bonawitz, Vincent D Costa, and R Becket Ebitz. Balancing exploration and exploitation with information and randomization. Current opinion in behavioral sciences, 38:49–56, 2021

    2021

  25. [33]

    Self-supervised graph learning for recommendation

    Jiancan Wu, Xiang Wang, Fuli Feng, Xiangnan He, Liang Chen, Jianxun Lian, and Xing Xie. Self-supervised graph learning for recommendation. In Proceedings of the 44th international ACM SIGIR conference on research and development in information retrieval, pages 726–735, 2021

    2021

  26. [34]

    Towards open-world recommendation with knowledge augmentation from large language models

    Yunjia Xi, Weiwen Liu, Jianghao Lin, Xiaoling Cai, Hong Zhu, Jieming Zhu, Bo Chen, Ruiming Tang, Weinan Zhang, and Yong Yu. Towards open-world recommendation with knowledge augmentation from large language models. In Proceedings of the 18th ACM Conference on Recommender System...

    2024

  27. [35]

    Hyper- graph contrastive collaborative filtering

    Lianghao Xia, Chao Huang, Yong Xu, Jiashu Zhao, Dawei Yin, and Jimmy Huang. Hyper- graph contrastive collaborative filtering. In Proceedings of the 45th International ACM SIGIR conference on research and development in information retrieval, pages 70–79, 2022

    2022

  28. [36]

    Hicf: Hyperbolic informative collaborative filtering

    Menglin Yang, Zhihao Li, Min Zhou, Jiahong Liu, and Irwin King. Hicf: Hyperbolic informative collaborative filtering. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 2212–2221, 2022

    2022

  29. [37]

    Hrcf: Enhancing col- laborative filtering via hyperbolic geometric regularization

    Menglin Yang, Min Zhou, Jiahong Liu, Defu Lian, and Irwin King. Hrcf: Enhancing col- laborative filtering via hyperbolic geometric regularization. In Proceedings of the ACM Web Conference 2022, pages 2462–2471, 2022

    2022

  30. [38]

    κhgcn: Tree-likeness modeling via con- tinuous and discrete curvature learning

    Menglin Yang, Min Zhou, Lujia Pan, and Irwin King. κhgcn: Tree-likeness modeling via con- tinuous and discrete curvature learning. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 2965–2977, 2023

    2023

  31. [39]

    Hyperbolic fine-tuning for large language models

    Menglin Yang, Aosong Feng, Bo Xiong, Jihong Liu, Irwin King, and Rex Ying. Hyperbolic fine-tuning for large language models. arXiv preprint arXiv:2410.04010, 2024

    2024

  32. [40]

    Hgformer: Hyperbolic graph transformer for recommendation

    Xin Yang, Xingrun Li, Heng Chang, Jinze Yang, Xihong Yang, Shengyu Tao, Ningkang Chang, Maiko Shigeno, Junfeng Wang, Dawei Yin, et al. Hgformer: Hyperbolic graph transformer for recommendation. arXiv preprint arXiv:2502.15693, 2024

    2024 arXiv

  33. [41]

    Graph convolutional neural networks for web-scale recommender systems

    Rex Ying, Ruining He, Kaifeng Chen, Pong Eksombatchai, William L Hamilton, and Jure Leskovec. Graph convolutional neural networks for web-scale recommender systems. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining, pages 974–983, 2018

    2018

  34. [42]

    Are graph augmentations necessary? simple graph contrastive learning for recommendation

    Junliang Yu, Hongzhi Yin, Xin Xia, Tong Chen, Lizhen Cui, and Quoc Viet Hung Nguyen. Are graph augmentations necessary? simple graph contrastive learning for recommendation. In Proceedings of the 45th international ACM SIGIR conference on research and development in informatio...

    2022

  35. [43]

    Avoiding monotony: improving the diversity of recommendation lists

    Mi Zhang and Neil Hurley. Avoiding monotony: improving the diversity of recommendation lists. In Proceedings of the 2008 ACM conference on Recommender systems, pages 123–130, 2008

    2008

  36. [44]

    Bertscore: Evaluating text generation with bert

    Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675, 2019

    1904 arXiv

  37. [45]

    Collm: Integrating collaborative embeddings into large language models for recommendation

    Yang Zhang, Fuli Feng, Jizhi Zhang, Keqin Bao, Qifan Wang, and Xiangnan He. Collm: Integrating collaborative embeddings into large language models for recommendation. arXiv preprint arXiv:2310.19488, 2023

    2023 arXiv

  38. [46]

    Lorentzian graph convolu- tional networks

    Yiding Zhang, Xiao Wang, Chuan Shi, Nian Liu, and Guojie Song. Lorentzian graph convolu- tional networks. In Proceedings of the web conference 2021, pages 1249–1261, 2021

    2021

  39. [47]

    Adapting large language models by integrating collaborative semantics for recommen- dation

    Bowen Zheng, Yupeng Hou, Hongyu Lu, Yu Chen, Wayne Xin Zhao, Ming Chen, and Ji-Rong Wen. Adapting large language models by integrating collaborative semantics for recommen- dation. In 2024 IEEE 40th International Conference on Data Engineering (ICDE) , pages 1435–1448. IEEE, 2024

    2024

  40. [48]

    Hierarchical collaborative embedding for context-aware recommendations

    Lei Zheng, Bokai Cao, Vahid Noroozi, S Yu Philip, and Nianzu Ma. Hierarchical collaborative embedding for context-aware recommendations. In 2017 IEEE International Conference on Big Data (Big Data), pages 867–876. IEEE, 2017

    2017

  41. [49]

    Improving recommendation lists through topic diversification

    Cai-Nicolas Ziegler, Sean M McNee, Joseph A Konstan, and Georg Lausen. Improving recommendation lists through topic diversification. In Proceedings of the 14th international conference on World Wide Web, pages 22–32, 2005. 12 A Proof In this section, we provide a rigorous math...

    2005

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.