Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

Beyond Semantic Understanding: Preserving Collaborative Frequency Components in LLM-based Recommendation

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Low-pass filters stop LLM recommenders from erasing collaborative signals by counteracting a layer-by-layer spectral decay the paper calls Intra-Layer Spectral Attenuation.

desk verdict The spectral attenuation result is a real, useful empirical finding, but the 'theoretically guaranteed' claim on TFM is not supported by the appendix proof. read the letter →

arxiv 2508.10312 v2 pith:63D3SDIH submitted 2025-08-14 cs.CL

Minhao Wang , Yunhang He , Cong Xu , Zhangchi Zhu , Shuang Hao , Ning Liu , Wei Zhang This is my paper
classification cs.CL
keywords sequentialrecommendationlargelanguagemodelscollaborativefilteringgraphsignalprocessingspectralattenuationfrequency-domaintemporalfrequencymodulationlow-pass
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 LLM-based recommenders systematically weaken collaborative signals as item embeddings pass through the transformer layers, even when the input contains pretrained collaborative ID embeddings. This happens because LLMs over-rely on semantic knowledge and implicitly suppress the low-frequency components of the item co-occurrence graph that carry collaborative filtering information. The paper proposes FreLLM4Rec, which purifies input embeddings with a global graph low-pass filter and then applies a temporal frequency modulation after each LLM layer to actively preserve these low-frequency components. The preservation mechanism is supported by a theorem connecting temporal low-pass filtering to graph spectral smoothness, and experiments on four datasets report NDCG@10 improvements of up to 8% over the best baseline.

What carries the argument

The load-bearing identity is the equality between the Graph Fourier Transform on a ring graph and the Discrete Fourier Transform, with eigenvalues $\lambda_k = 2 - 2\cos(2\pi k/T)$. This lets the paper replace expensive, non-parallelizable local graph filtering with a one-dimensional FFT-based Butterworth low-pass filter along the sequence dimension. Under the Spatio-Temporal Locality assumption, temporal smoothing reduces the graph Laplacian quadratic form and shifts energy toward low graph frequencies, which is exactly the collaborative signal the paper wants to preserve.

What would settle it

On a dataset where temporally adjacent interactions are deliberately anti-correlated (e.g., users alternate between two unrelated item clusters), compute the graph Laplacian quadratic form before and after a temporal low-pass filter; if the form does not decrease, Theorem 1's premise fails. Alternatively, measure the alignment between the DFT basis and the eigenvectors of the local co-occurrence graph Laplacian across real user subsequences; low alignment would break the claimed equivalence.

Watch

Extended reading notes

Core claim

The paper claims that LLM-based recommenders exhibit Intra-Layer Spectral Attenuation: as embeddings propagate through the LLM backbone, the energy of low-frequency collaborative signals—the smooth, community-level patterns that encode item similarity and user preference—monotonically decreases. This is contrasted with a purpose-built transformer sequential recommender, which preserves or even enhances these signals. The paper then introduces FreLLM4Rec, which combines a Global Graph Low-Pass Filter on the input item embeddings with Temporal Frequency Modulation applied after each LLM layer. The key theoretical claim is that under a Spatio-Temporal Locality assumption, temporal low-pass filt

Load-bearing premise

The guarantee rests on Assumption 1: items that appear close together in a user's timeline are more collaboratively similar than temporally distant items, so smoothing along time is equivalent to smoothing along the co-occurrence graph; if that locality fails, temporal filtering will not preserve collaborative low-frequency components.

Editorial extensions

If this is right

  • If the attenuation claim is right, LLM-based recommenders that only fuse semantic and collaborative embeddings at the input will keep underperforming; corrective spectral layers are needed after each backbone layer.
  • Because the correction is architecture-agnostic, the same G-LPF plus TFM recipe should transfer to other LLM backbones and other sources of collaborative embeddings, which the paper tests across three backbones and three embedding sources.
  • The layer-wise spectral energy analysis provides a diagnostic that can be applied to any sequence model to detect where collaborative information is lost.
  • The reported gains suggest the bottleneck in LLM recommendation is not semantic understanding alone but the preservation of collaborative structure during propagation.

Reading between the lines

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

  • Left implicit: the same spectral audit could be run on other structured-input LLM tasks, such as time-series forecasting or trajectory modeling, where low-frequency relational signal matters; the attenuation may be a general property of LLM pre-training rather than recommendation-specific.
  • A testable extension would be an adaptive cutoff frequency learned per layer; the paper's hyperparameter sensitivity analysis indicates the optimal cutoff varies by dataset and depth, so per-layer adaptation could improve on the fixed setting.
  • One could probe causality by applying TFM only at the last layer: if attenuation is progressive, last-layer-only correction should help less than per-layer correction, which the paper's energy plots imply but do not test directly.
  • The theoretical bridge assumes the DFT basis approximates the eigenvectors of the local co-occurrence graph Laplacian; measuring their alignment on real user subsequences would directly quantify how often the assumption holds.
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 / 6 minor

Summary. The paper studies how pretrained LLM backbones process collaborative ID embeddings for sequential recommendation. It reports a layer-wise spectral analysis of hidden states and identifies 'Intra-Layer Spectral Attenuation': low-frequency components of the item co-occurrence graph decay across LLM layers, unlike in SASRec. To counteract this, the authors propose FreLLM4Rec, which combines a global graph low-pass filter (G-LPF) on input ID embeddings with a temporal Butterworth low-pass filter (TFM) applied after each LLM layer. They claim a theoretical guarantee that TFM preserves collaborative low-frequency graph components via a connection between temporal DFT filtering and graph spectral smoothing. Experiments on four datasets report consistent NDCG@10/Recall@10 gains, up to 8% over the best baseline, with ablations and cross-backbone/cross-embedding robustness.

Significance. The paper addresses an important and underexplored issue: how LLM-based recommenders transform collaborative signals internally. The proposed phenomenon, if validated, is useful for understanding LLM-recommendation integration, and the method is simple, efficient, and shows consistent improvements across multiple datasets, backbones, and embedding sources. The release of code and the breadth of ablations are strengths. However, the paper's central theoretical claim (Theorem 1) is not established by the provided proof, and the dimensional inconsistency in the local spectral analysis Definition 1 undermines the measurement protocol used to demonstrate the phenomenon. The empirical results are promising, but the current manuscript overstates the theoretical foundations.

major comments (4)
  1. [Appendix D / Theorem 1] The proof of Theorem 1 is not valid as written. Part 1 asserts that a temporal low-pass filter acts as a local averaging function and invokes Jensen's inequality to conclude Eq. (13). But the implemented Butterworth filter in Eq. (8)-(9) is not a positive local averaging kernel; its impulse response has negative lobes and can increase adjacent differences, so variance reduction and Eq. (13) do not follow without additional bounds that are not derived. Part 2 argues from Eq. (15) and approximate energy conservation (Eq. (16)) that energy must shift to low graph frequencies. A decrease in the frequency-weighted sum does not imply concentration in the lowest bands—mass could move from mid-frequencies to both low and high frequencies—and a low-pass filter generally reduces total energy, so Eq. (16) is unjustified unless renormalization is explicitly defined. The abstract's claim that 'the co
  2. [Definition 1 / Eq. (4)] The local spectral analysis is dimensionally inconsistent. Definition 1 defines A_u ∈ R^{T×T} (T nodes), but the layer-wise hidden states are H_u^l ∈ R^{(T-1)×d}. The graph Fourier transform U_u^T H_u^l is undefined because the number of nodes in A_u does not match the number of rows of H_u^l. This measurement is the basis for the central empirical claim of Intra-Layer Spectral Attenuation in Figures 1 and 8, so it needs to be re-derived with a consistent graph dimension (e.g., (T-1)×(T-1)) or with a clear explanation of how A_u is constructed for the T-1 target items.
  3. [Table 2 / §5.2] The central empirical claim rests on single numbers without error bars or significance tests. Given the reported gains of 4-8% over the best baseline, it is important to report mean ± std over multiple seeds (at least 3) and to state whether the improvements are statistically significant. Without this, the reader cannot assess whether the observed gains are robust to randomness in the LLM pipeline.
  4. [Assumption 1 / Appendix C] The theoretical connection between temporal and graph frequency domains depends on Assumption 1 (spatio-temporal locality) and on treating user sequences as ring graphs. Neither is quantitatively validated in the manuscript. The Appendices cite support from prior work, but the specific datasets used here may not satisfy the assumption uniformly; a simple check (e.g., average co-occurrence edge weight as a function of temporal lag) would strengthen the claim. As written, the proof does not address boundary effects or non-ring sequence structure, so the guarantee holds only under conditions that are not established.
minor comments (6)
  1. [Definition 1] The notation T is used inconsistently: S_u is defined as (v_1,...,v_{T-1}), but A_u is said to be T×T. Please align the indexing throughout the definition.
  2. [Eq. (4)] The energy expression \|(\hat H_u^l)_k\|_F^2 is ambiguous. Clarify whether the k-th frequency component is a vector over embedding dimensions and how the Frobenius norm is taken.
  3. [Figures 1 and 8] The spectral energy plots would benefit from axis labels, units, and a description of how the 'low-frequency band' is partitioned in the quantile-based analysis.
  4. [§4.2 / Eq. (9)] The Butterworth filter order n is not reported in the hyperparameter discussion in Appendix B (only α and ω_c are tuned). Please report the value or range used.
  5. [Table 2] The LLaMA-3 baseline appears to be evaluated in a zero-shot setting, whereas the other LLM baselines use the embedding-as-token paradigm. This makes the comparison potentially unfair; state explicitly how LLaMA-3 was adapted.
  6. [§3.2] The text says low-frequency energy decays 'exponentially' and 'over 70%' is lost, but no numerical quantification or fitted curve is provided. Please support these statements with concrete numbers.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the empirical pipeline is self-contained and the theoretical claim, while close to the definition of low-pass filtering, is not a fitted input or a self-citation chain.

full rationale

The paper's central empirical result (up to 8.00% NDCG@10 improvement over the best baseline) is obtained by training and evaluation on standard benchmarks against external baselines, with hyperparameters tuned on validation sets, so the performance claim is not a fitted input renamed as a prediction. The discovery of Intra-Layer Spectral Attenuation is an independent measurement using local graph spectral energy and is replicated across multiple LLM backbones and datasets. The theoretical claim in Theorem 1 / Appendix D is not circular in the prohibited sense: it is conditional on Assumption 1, an external spatio-temporal locality premise, and the proof attempts to derive the graph-smoothness reduction rather than assuming the conclusion. The proof does contain mathematically unsupported steps—e.g., the assertion that a temporal low-pass filter acts as a local averaging operator and that Eq. (13) follows via Jensen, and the non-sequitur in Part 2 from Eq. (15)–(16) to concentration in the lowest bands—but invalidity is a correctness risk, not a circular reduction. The only self-citation ([65], Zhu & Zhang 2025) appears alongside multiple independent references for the standard low-frequency premise and is not load-bearing for the central argument. Overall circularity score: 0.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central method depends on two tunable filtering hyperparameters (alpha and omega_c) and a third unspecified order n, plus a domain assumption (Spatio-Temporal Locality) that is not empirically validated in this paper. No genuinely new entities are introduced.

free parameters (3)
  • alpha (G-LPF strength) = tuned in [0, 1.0], optimal around 0.5
    Controls the first-order low-pass filter h(lambda) = 1 - alpha*lambda applied to item embeddings. Tuned on validation data (Appendix B, Figure 5).
  • omega_c (TFM cutoff frequency) = tuned in [0, 1.0], optimal range 0.1-0.5
    Cutoff frequency of the Butterworth filter applied after each LLM layer. Tuned on validation data (Appendix B, Figure 6).
  • Butterworth filter order n = not reported
    Order of the Butterworth filter in Equation (9). Neither the value nor the tuning procedure is specified in the paper.
assumptions (3)
  • domain assumption Assumption 1 (Spatio-Temporal Locality): items appearing in temporal proximity within user sequences exhibit higher collaborative similarity than temporally distant items.
    Introduced in Section 4.2 and used to prove Theorem 1 in Appendix D. If false, temporal smoothing does not necessarily increase graph smoothness.
  • standard math The DFT is equivalent to the Graph Fourier Transform on a ring graph, with eigenvalues lambda_k = 2 - 2*cos(2*pi*k/T).
    Invoked in Appendix C to connect temporal frequency filtering to graph spectral filtering. This holds for ring graphs but is an approximation for real user sequences.
  • domain assumption Collaborative information in item embeddings predominantly lies in low-frequency components of the item co-occurrence graph.
    Used throughout the paper (Sections 1 and 2.2) and attributed to prior graph-based collaborative filtering literature [13, 46, 52, 65].

how reviews work

0 comments
Cite this review

Pith. "Pith review of Beyond Semantic Understanding: Preserving Collaborative Frequency Components in LLM-based Recommendation." pith.science (2026). https://pith.science/paper/63D3SDIH

@misc{pith2026250810312,
  author       = {Pith},
  title        = {Pith review of: Beyond Semantic Understanding: Preserving Collaborative Frequency Components in LLM-based Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/63D3SDIH}},
  note         = {Machine review of arXiv:2508.10312}
}
read the original abstract

Recommender systems in concert with Large Language Models (LLMs) present promising avenues for generating semantically-informed recommendations. However, LLM-based recommenders exhibit a tendency to overemphasize semantic correlations within users' interaction history. When taking pretrained collaborative ID embeddings as input, LLM-based recommenders progressively weaken the inherent collaborative signals as the embeddings propagate through LLM backbones layer by layer, as opposed to traditional Transformer-based sequential models in which collaborative signals are typically preserved or even enhanced for state-of-the-art performance. To address this limitation, we introduce FreLLM4Rec, an approach designed to balance semantic and collaborative information from a spectral perspective. Item embeddings that incorporate both semantic and collaborative information are first purified using a Global Graph Low-Pass Filter (G-LPF) to preliminarily remove irrelevant high-frequency noise. Temporal Frequency Modulation (TFM) then actively preserves collaborative signal layer by layer. Note that the collaborative preservation capability of TFM is theoretically guaranteed by establishing a connection between the optimal but hard-to-implement local graph fourier filters and the suboptimal yet computationally efficient frequency-domain filters. Extensive experiments on four benchmark datasets demonstrate that FreLLM4Rec successfully mitigates collaborative signal attenuation and achieves competitive performance, with improvements of up to 8.00\% in NDCG@10 over the best baseline. Our findings provide insights into how LLMs process collaborative information and offer a principled approach for improving LLM-based recommendation systems.

Figures

Figures reproduced from arXiv: 2508.10312 by the authors.

Figure 1
Figure 1. Discovery of the Intra-Layer Spectral Attenuation [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of FreLLM4Rec: Our frequency-aware ap [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Impact of retaining percentages of low-frequency [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Empirical validation of spectral decay mitigation. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 7
Figure 7. Figure 7: A ring graph with N nodes. C Connecting Temporal and Graph Frequency Domains The Discrete Fourier Transform (DFT) for a temporal signal x = [𝑥0, 𝑥1, . . . , 𝑥𝑇 −1] 𝑇 is defined as: F (x) [𝑘] = 𝑇∑︁−1 𝑡=0 𝑥𝑡𝑒 −𝑗2𝜋𝑘𝑡/𝑇 , (10) where 𝑘 = 0, 1, . . . ,𝑇 − 1 represents the fr…
Figure 5
Figure 5. Figure 5: G-LPF filtering strength analysis. Moderate filtering [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: TFM cutoff frequency analysis. Lower cutoffs better [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 8
Figure 8. Figure 8: Spectral attenuation phenomenon across different LLMs on All Beauty and Luxury Beauty datasets [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Private Approximation of Graph Spectra and Cuts via Spectral Amplifiers

    cs.DS 2026-07 conditional novelty 8.0 of 10

    An edge-differentially-private algorithm approximates every cut of any unweighted n-vertex graph with error γ·cut + Õ(n^(13/12+o(1))), beating the previous polynomial-time bound of n^(5/4+o(1)).

Reference graph

Works this paper leans on

66 extracted references · 38 canonical work pages · cited by 1 Pith paper

  1. [1]

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

    2023

  2. [2]

    2004.Convex optimization

    Stephen Boyd and Lieven Vandenberghe. 2004.Convex optimization. Cambridge university press

    2004

  3. [3]

    Iván Cantador, Peter Brusilovsky, and Tsvi Kuflik. 2011. Second workshop on information heterogeneity and fusion in recommender systems (HetRec2011). In Proceedings of the fifth ACM conference on Recommender systems. 387–388

    2011

  4. [4]

    Yang Chen, Yixuan Wang, Ting Dong, and Xiangnan He. 2025. Text-like Encoding of Collaborative Information in Large Language Models for Recommendation. In Findings of the Association for Computational Linguistics: ACL 2025. 3456–3470

    work page 2025

  5. [5]

    Zhiqian Chen, Feng Chen, Lei Zhang, Taoran Ji, Kaidi Fu, Liang Zhao, Fanglan Chen, Lingfei Wu, Charu Aggarwal, and Chang-Tien Lu. 2022. Revisiting the role of similarity based graph filtering in graph neural networks. InProceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 121–131

    work page 2022

  6. [6]

    Michaël Defferrard, Xavier Bresson, and Pierre Vandergheynst. 2016. Convolu- tional neural networks on graphs with fast localized spectral filtering. InAdvances in neural information processing systems, Vol. 29

    work page 2016

  7. [7]

    Xinyu Du, Huanhuan Yuan, Pengpeng Zhao, Jianfeng Qu, Fuzhen Zhuang, Guan- feng Liu, Yanchi Liu, and Victor S Sheng. 2023. Frequency enhanced hybrid attention network for sequential recommendation.Proceedings of the 46th In- ternational ACM SIGIR Conference on Research and Development in Information Retrieval(2023), 78–88

    work page 2023

  8. [8]

    Wenqi Fan, Zihuai Zhao, Jiatong Li, Yunqing Liu, Xiaowei Mei, Yiqi Wang, Jiliang Tang, and Qing Li. 2023. Recommender systems in the era of large language models (LLMs).arXiv preprint arXiv:2307.02046(2023)

    arXiv 2023

Show all 66 references
  1. [9]

    Zihan Gao, Qingquan Qiu, Yang Liu, and Min Zhang. 2024. Parameter-efficient fine-tuning of large language models through Fourier transforms. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 1234–1248

    2024

  2. [10]

    Shijie Geng, Shuchang Liu, Zuohui Fu, Yingqiang Ge, and Yongfeng Zhang. 2022. Recommendation as language processing (RLP): A unified pretrain, personalized prompt & predict paradigm (P5). InProceedings of the 16th ACM Conference on Recommender Systems. 299–315

    2022

  3. [11]

    Jesse Harte, Wouter Zorgdrager, Panos Louridas, Asterios Katsifodimos, Diet- mar Jannach, and Marios Fragkoulis. 2024. Leveraging Large Language Models for Sequential Recommendation. InProceedings of the 18th ACM Conference on Recommender Systems. 1096–1108

    2024

  4. [12]

    Ruining He and Julian McAuley. 2016. Fusing similarity models with markov chains for sparse sequential recommendation. In2016 IEEE 16th international conference on data mining (ICDM). IEEE, 191–200

    2016

  5. [13]

    Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yongdong Zhang, and Meng Wang. 2020. LightGCN: Simplifying and powering graph convolution network for recommendation. InProceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval. 639–648

    2020

  6. [14]

    Xiangnan He, Lizi Liao, Hanwang Zhang, Liqiang Nie, Xia Hu, and Tat-Seng Chua. 2017. Neural collaborative filtering. InProceedings of the 26th international conference on world wide web. 173–182

    2017

  7. [15]

    Yingzhi He, Xiaohao Liu, An Zhang, Yunshan Ma, and Tat-Seng Chua. 2025. LLM2Rec: Large Language Models Are Powerful Embedding Models for Sequen- tial Recommendation. arXiv:2506.21579 [cs.IR] https://arxiv.org/abs/2506.21579

    2025 arXiv

  8. [16]

    Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk

  9. [17]

    Yupeng Hou, Shanlei Li, Zhoujun Liu, Xiaopeng Yao, Yuan Yao, and Jinyang Huai. 2024. E4SRec: An Elegant Effective Efficient Extensible Solution for Large Language Model Based Sequential Recommendation. InProceedings of the 47th International ACM SIGIR Conference on Research an...

    2024

  10. [18]

    Yupeng Hou, Junjie Zhang, Zihan Lin, Hongyu Lu, Ruobing Xie, Julian McAuley, and Wayne Xin Zhao. 2024. Large language models are zero-shot rankers for recommender systems. InEuropean Conference on Information Retrieval. Springer, 364–381

    2024

  11. [19]

    Panfeng Ji, Shuai Wang, Yupeng Hou, and Fuli Feng. 2024. GenRec: Generative Sequential Recommendation with Large Language Models. InAdvances in Neural Information Processing Systems, Vol. 37

    2024

  12. [20]

    Wang-Cheng Kang and Julian McAuley. 2018. Self-attentive sequential recom- mendation. In2018 IEEE International Conference on Data Mining (ICDM). IEEE, 197–206

    2018

  13. [21]

    Hongyu Li, Mao Zhang, Fei Chen, and Xin Xia. 2024. LLM-Enhanced Cross- Domain Sequential Recommendation. InProceedings of the 33rd ACM International Conference on Information and Knowledge Management. 2317–2326

    2024

  14. [22]

    Lei Li, Yongfeng Zhang, and Li Chen. 2024. Large Language Models for Generative Recommendation: A Survey and Visionary Discussions. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 4321–4331

    2024

  15. [23]

    Jiayi Liao, Sihang Li, Zhengyi Yang, Jiancan Wu, Yancheng Yuan, Xiang Wang, and Xiangnan He. 2024. LLaRA: Large Language-Recommendation Assistant. InProceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 2132–2136

    2024

  16. [24]

    Bin Liu, Tao Zhu, Jianxun Pan, and Fuli Feng. 2025. Triple Modality Fusion: Aligning Visual, Textual, and Graph Data with Large Language Models for Multi- Behavior Recommendations. InProceedings of the ACL 2025 Conference. 789–803

    2025

  17. [25]

    Peng Liu, Lemei Zhang, and Jon Atle Gulla. 2023. Pre-train, prompt and recom- mendation: A comprehensive survey of language modelling paradigm adapta- tions in recommender systems.Transactions of the Association for Computational Linguistics11 (2023), 1553–1571

    2023

  18. [26]

    Julian McAuley and Jure Leskovec. 2013. Hidden factors and hidden topics: understanding rating dimensions with review text. InProceedings of the 7th ACM conference on Recommender systems. 165–172

    2013

  19. [27]

    Jianmo Ni, Jiacheng Li, and Julian McAuley. 2019. Justifying recommendations using distantly-labeled reviews and fine-grained aspects. InProceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural...

    2019

  20. [28]

    Hoang NT and Takanori Maehara. 2019. Revisiting graph neural networks: All we have is low-pass filters. InarXiv preprint arXiv:1905.09550

    2019 arXiv

  21. [29]

    Antonio Ortega, Pascal Frossard, Jelena Kovačević, José MF Moura, and Pierre Vandergheynst. 2018. Graph signal processing: Overview, challenges, and appli- cations.Proc. IEEE106, 5 (2018), 808–828

    2018

  22. [30]

    Xubin Ren, Wei Wei, Lianghao Xia, Lixin Su, Suqi Cheng, Junfeng Wang, Dawei Yin, and Chao Huang. 2024. Representation Learning with Large Language Models for Recommendation. InProceedings of the ACM Web Conference 2024. 3464–3475

    2024

  23. [31]

    Zhaopeng Ren, Zihan Wang, Zijian Ke, Fuzhen Zhuang, Zhenya Liu, Qing Zhu, and Qi Liu. 2025. Large Language Models meet Collaborative Filtering: An Efficient All-round LLM-based Recommender System. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data ...

    2025

  24. [32]

    Steffen Rendle, Christoph Freudenthaler, and Lars Schmidt-Thieme. 2010. Factor- izing personalized markov chains for next-basket recommendation. InProceedings of the 19th international conference on World wide web. 811–820

    2010

  25. [33]

    Yehjin Shin, Jeongwhan Choi, Hyowon Wi, and Noseong Park. 2024. An Atten- tive Inductive Bias for Sequential Recommendation beyond the Self-Attention. arXiv:2312.10325 [cs.LG] https://arxiv.org/abs/2312.10325

    2024 arXiv

  26. [34]

    David I Shuman, Sunil K Narang, Pascal Frossard, Antonio Ortega, and Pierre Vandergheynst. 2013. The emerging field of signal processing on graphs: Ex- tending high-dimensional data analysis to networks and other irregular domains. IEEE Signal Processing Magazine30, 3 (2013), 83–98

    2013

  27. [35]

    Fei Sun, Jun Liu, Jian-Yun Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang. 2019. BERT4Rec: Sequential recommendation with bidirectional encoder representations from transformer. InProceedings of the 28th ACM International Conference on Information and Knowledge Managemen...

    2019

  28. [36]

    Zhongxiang Sun, Zihua Si, Xiaoxue Zang, Kai Zheng, Yang Song, Xiao Zhang, and Jun Xu. 2024. Large Language Models Enhanced Collaborative Filtering. arXiv preprint arXiv:2403.17688(2024)

    2024 arXiv

  29. [37]

    Youhua Tan, Chang Liu, Zhiqi Wei, Guannan Chen, and Xiaochun Zheng. 2024. ID vs. Modality: Which is Better for Sequential Recommendation?. InProceedings of the 33rd ACM International Conference on Information and Knowledge Management. 3421–3430. Conference acronym ’XX, June 03...

    2024

  30. [38]

    Jiaxi Tang and Ke Wang. 2018. Personalized Top-N Sequential Recommendation via Convolutional Sequence Embedding. arXiv:1809.07426 [cs.IR] https://arxiv. org/abs/1809.07426

    2018 arXiv

  31. [39]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. LLaMA: Open and efficient foundation language models.arXiv preprint arXiv:2302.13971(2023)

    2023 arXiv

  32. [40]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. InAdvances in neural information processing systems, Vol. 30

    2017

  33. [41]

    Lingfeng Wang, Jiayu Zhang, Xiang Chen, Yujie Ma, and Defu Lian. 2025. Order- agnostic Identifier for Large Language Model-based Generative Recommendation. InProceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval. 123–132

    2025

  34. [42]

    Shoujin Wang, Liang Hu, Yan Wang, Longbing Cao, Quan Z Sheng, and Mehmet Orgun. 2021. A survey on deep learning for recommender systems.Knowledge and Information Systems61, 1 (2021), 1–50

    2021

  35. [43]

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

    2019

  36. [44]

    Wei Wei, Xubin Ren, Jiabin Tang, Qinyong Wang, Lixin Su, Suqi Cheng, Junfeng Wang, Dawei Yin, and Chao Huang. 2024. LLMRec: Large Language Models with Graph Augmentation for Recommendation. InProceedings of the 18th ACM International Conference on Web Search and Data Mining. 806–815

    2024

  37. [45]

    Felix Wu, Amauri Souza, Tianyi Zhang, Christopher Fifty, Tao Yu, and Kilian Weinberger. 2019. Simplifying graph convolutional networks. InInternational conference on machine learning. PMLR, 6861–6871

    2019

  38. [46]

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

    2021

  39. [47]

    Likang Wu, Zhi Zheng, Zhaopeng Qiu, Hao Wang, Hongchao Gu, Tingjia Shen, Chuan Qin, Chen Zhu, Hengshu Zhu, Qi Liu, et al . 2023. A survey on large language models for recommendation.arXiv preprint arXiv:2305.19860(2023)

    2023 arXiv

  40. [48]

    Shu Wu, Yuyuan Tang, Yanqiao Zhu, Liang Wang, Xing Xie, and Tieniu Tan. 2019. Session-based recommendation with graph neural networks. InProceedings of the AAAI conference on artificial intelligence, Vol. 33. 346–353

    2019

  41. [49]

    Xu Xie, Fei Sun, Zhaoyang Liu, Shiwen Wu, Jinyang Gao, Jiandong Zhang, Bolin Ding, and Bin Cui. 2022. Contrastive learning for sequential recommendation. In 2022 IEEE 38th international conference on data engineering (ICDE). IEEE, 1259– 1273

    2022

  42. [50]

    An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Cheng- peng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. 2024. Qwen2 Technical Report.arXiv preprint arXiv:2407.10671(2024)

    2024 arXiv

  43. [51]

    Yaowen Ye, Lianghao Xia, and Chao Huang. 2023. Graph Masked Autoencoder for Sequential Recommendation. InProceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR ’23). ACM, 321–330. doi:10.1145/3539618.3591692

    2023

  44. [52]

    Junliang Yu, Xin Xia, Tong Chen, Lizhen Cui, Nguyen Quoc Viet Hung, and Hongzhi Yin. 2023. XSimGCL: Towards extremely simple graph contrastive learning for recommendation. InIEEE Transactions on Knowledge and Data Engi- neering, Vol. 36. 913–926

    2023

  45. [53]

    Lu Yu, Jun Xie, and Xiang Zheng. 2023. The relationship between graph Fourier transform (GFT) and discrete cosine transform (DCT) for 1D signal and 2D image. Signal, Image and Video Processing17, 2 (2023), 445–451

    2023

  46. [54]

    Haipeng Yuan, Shuai Wang, Zhen Li, and Xiangnan He. 2024. Cross-Domain Sequential Recommendation via Neural Process. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 38. 9876–9884

    2024

  47. [55]

    Xiaoyu Yuan, Yang Liu, Xin Xia, and Jiliang Yin. 2024. Customizing Language Models with Instance-wise LoRA for Sequential Recommendation. InAdvances in Neural Information Processing Systems, Vol. 38

    2024

  48. [56]

    Zheng Yuan, Fajie Yuan, Yu Song, Youhua Li, Junchen Fu, Fei Yang, Yunzhu Pan, and Yongxin Ni. 2023. Where to Go Next for Recommender Systems? ID- vs. Modality-based Recommender Models Revisited. InProceedings of the 46th International ACM SIGIR Conference on Research and Devel...

    2023

  49. [57]

    Jiaqi Zhai, Lucy Liao, Xing Liu, Yueming Wang, Rui Li, Xuan Cao, Leon Gao, Zhao- jie Gong, Fangda Gu, Michael He, et al. 2024. Actions speak louder than words: Trillion-parameter sequential transducers for generative recommendations.arXiv preprint arXiv:2402.17152(2024)

    2024 arXiv

  50. [58]

    Juntao Zhang, Yiqun Xie, Hao Zhang, Menghan Yu, Shuguang Chen, Bo Shen, Menghan Zhang, Weixiang Chen, Chi-min Wu, and Peng Cui. 2024. IDGenRec: LLM-RecSys Alignment with Textual ID Learning. InProceedings of the 47th International ACM SIGIR Conference on Research and Developme...

    2024

  51. [59]

    Zeyu Zhang, Xiaocan Feng, Yong Wang, and Defu Lian. 2024. Ranking-based Prompting for Large Language Models in Recommendation. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 8901–8915

    2024

  52. [60]

    Tianyu Zhao, Mingyu Li, Xin Wang, and Zhuo Chen. 2024. Spectral filters for efficient fine-tuning of large language models. InProceedings of the 41st International Conference on Machine Learning. 8901–8915

    2024

  53. [61]

    Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. 2023. A survey of large language models.arXiv preprint arXiv:2303.18223(2023)

    2023 arXiv

  54. [62]

    Xinyu Zhao, Hansi Ren, Lei Li, and Yue Wang. 2024. Lifelong Sequential Rec- ommendation with Large Language Models. InProceedings of the International Conference on Learning Representations

    2024

  55. [63]

    Kun Zhou, Hui Wang, Wayne Xin Zhao, Yutao Zhu, Sirui Wang, Fuzheng Zhang, Zhongyuan Wang, and Ji-Rong Wen. 2020. S3-rec: Self-supervised learning for se- quential recommendation with mutual information maximization. InProceedings of the 29th ACM international conference on inf...

    2020

  56. [64]

    Kun Zhou, Hui Yu, Wayne Xin Zhao, and Ji-Rong Wen. 2022. Filter-enhanced MLP is all you need for sequential recommendation. InProceedings of the ACM Web Conference 2022. 2388–2399

    2022

  57. [65]

    Zhangchi Zhu and Wei Zhang. 2025. Exploring Feature-based Knowledge Dis- tillation for Recommender System: A Frequency Perspective. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 1. 2182–2193. A Baseline Details In this paper, we cons...

    2025

  58. [2015]

    InarXiv preprint arXiv:1511.06939

    Session-based recommendations with recurrent neural networks. InarXiv preprint arXiv:1511.06939

    arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.