Pith. sign in

REVIEW 5 major objections 5 minor 38 references

TRWH claims that combining LLM-generated user and item profiles with a heterogeneous graph network sharply improves rating prediction on sparse Amazon data—and that random-walk-added links help word-vector embeddings but wash out LLM nuance

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

TRWH combines LLM profiles, a heterogeneous GNN, and one-hop random walk edges, but its headline performance claims are contradicted by its own results table.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection The random-walk-vs-LLM-embedding finding is worth a look, but the headline results are not in the table and the evaluation likely leaks test reviews; desk reject. the 5 major comments →

arxiv 2607.25471 v1 pith:CBN42SWW submitted 2026-07-28 cs.AI cs.MM

TRWH: A Text-Driven Random Walk Heterogeneous GNN for Semantic-Aware Sparse Recommendation

He Ma , Chen Liu This is my paper
classification cs.AI cs.MM
keywords recommendationgraph neural networklarge language modelrandom walkheterogeneous graphsparse datarating predictiontext embedding
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 is trying to establish a practical recipe for sparse recommendation: use LLMs to write semantic profiles of users and items, feed those as node features into a heterogeneous graph network with multiple relationship types, and enrich the graph with one-hop random-walk links that connect similar users and items. The authors report large reductions in RMSE and MAE on two Amazon subsets, and they identify a counterintuitive interaction: the added random-walk edges improve traditional word embeddings but degrade the richer LLM embeddings, implying the augmentation should be applied selectively. If correct, the work tells engineers when graph augmentation helps, and it gives a way to combine text understanding with collaborative filtering without losing semantic detail.

Core claim

On its own terms, TRWH shows that fusing LLM-generated textual profiles with a heterogeneous graph—whose edge types include ratings, reviews, purchases, same-store links, and random-walk-inferred similarities—yields rating error reductions of up to 80% RMSE and 52.6% MAE on Amazon Fashion and 25.7% and 10.8% on Beauty relative to strong baselines. The paper's more consequential claim is the interaction effect: one-hop random-walk augmentation consistently improves Word2Vec-based models, but applying the same augmentation to LLM-based embeddings hurts both RMSE and MAE, because averaging over added neighbors dulls the fine-grained semantic distinctions the LLM encoded.

What carries the argument

The central mechanism is the one-hop random walk that generates second-order proximity edges: from a user, sample an item they rated, then sample another user who rated that same item, and add a bidirectional related-user edge; symmetrically for items. This creates user-user and item-item links that densify the sparse bipartite graph. These edges feed a heterogeneous graph network with a single propagation layer that exchanges information across all edge types, using LLM profile embeddings (or Word2Vec embeddings) as node features. The second key piece is the LLM profile generator, which uses structured prompts with reasoning steps to turn a user's reviews and item titles/descriptions into a

Load-bearing premise

The evaluation assumes the LLM never sees the review text of the items whose ratings are being predicted; if those reviews are included in the user profile prompts, the model is effectively reading the answer and the reported error reductions are not real.

What would settle it

Regenerate all user profiles per cross-validation fold, removing every test-fold review from the prompt text, and retrain TRWH with identical hyperparameters; if the RMSE/MAE advantage over word-vector baselines collapses (or the method no longer beats matrix factorization), the central performance claim fails. Also, a simple check: for a held-out user-item pair, verify whether the test review string appears in the profile used for prediction.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Sparse graphs can be enriched with second-order user-user and item-item edges from one-hop random walks, improving learning without the noise of longer walks.
  • Graph augmentation is not universally beneficial: it helps shallow embeddings and hurts semantically rich ones, so the augmentation strategy should depend on the embedding type.
  • LLM-generated profiles can serve as node features in a heterogeneous GNN, preserving semantic distinctiveness when the graph is not over-augmented.
  • Explicit modeling of multiple relationship types (ratings, reviews, purchases, store membership, inferred similarity) improves prediction over single-relational graph propagation.
  • On Fashion and Beauty, the proposed models achieve large RMSE/MAE improvements over the compared baselines.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The reported gains may be inflated by evaluation leakage: user profiles are generated from review texts without evidence that test-fold reviews are excluded, so a per-fold regenerated profile experiment is needed before taking the error reductions at face value.
  • The dilution effect suggests a testable design rule: measure an embedding's semantic sharpness (e.g., average distance to nearest neighbors) and apply random-walk edges only where sharpness is low; the paper's result predicts a crossover.
  • The same profile-generation machinery could be applied to cold-start users and items, where only text is available, though the paper did not test this scenario.
  • The improvements over cross-domain baselines may overstate real gains; a direct comparison against single-domain matrix factorization and graph models on identical splits would give a fairer picture.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper proposes TRWH, a framework for rating prediction that combines text-derived user/item embeddings with a heterogeneous GNN and random-walk augmentation. Embeddings are produced either by Word2Vec or by LLM-generated textual profiles (Llama-3.2-3B-Instruct + Instructor-XL). A heterogeneous graph with nine edge types (rating, review, purchase, same-store, and random-walk-based user/item similarity edges) is then used for message passing, and a prediction head outputs ratings. Experiments on Amazon 2023 Fashion and Beauty report large RMSE/MAE reductions over a set of baselines, and an ablation claims that random-walk augmentation helps Word2Vec embeddings but dilutes LLM embeddings. The paper contains an abstract, introduction, literature review, methodology, evaluation, experiment setup, conclusion, limitations, and a NeurIPS checklist.

Significance. If the empirical results were trustworthy, the paper would make a useful empirical contribution to integrating LLM-derived semantic profiles with heterogeneous GNNs for sparse recommendation, and the random-walk-versus-LLM-dilution insight is thought-provoking. The proposed architecture is clearly described and the datasets are large and realistic. However, the paper does not currently provide credible evidence for its central claims: the abstract's improvement percentages are not derivable from Table 2, the evaluation protocol appears to permit leakage of test-fold information into both the LLM profiles and the graph, no error bars or significance tests are reported, and most baselines are cross-domain methods not designed or tuned for single-domain rating prediction. No code or detailed prompts are supplied, and the manuscript contains unfinished placeholder text. For these reasons, the contribution cannot be assessed as presented.

major comments (5)
  1. [§3.1.2, Eq. (3); §5.3] The user prompt Qu in Eq. (3) concatenates cv=[α, θ, r^u_v] for every v∈Iu, where r^u_v is the user's own review text. The paper never states that profiles are regenerated per fold or that Iu excludes test items. With 5-fold CV, if profiles are built once from the full interaction history, the LLM embedding for a test user can encode the sentiment/review text of the very items whose ratings are predicted, and the HeteroGNN then propagates this along user–item edges. The prompt details are not provided ('provided in (create a figure)'), so the reader cannot verify any exclusion. This makes the reported RMSE/MAE measures of memorization rather than generalization.
  2. [Abstract; Table 2] The headline reductions are not derivable from Table 2. On Fashion, the best TRWH RMSE (LLMHet, 1.0604) is 7.8% better than the strongest baseline (Homogeneous GNN, 1.15), and the best MAE (W2VRHet, 0.9089) is 34.7% better than PEMF-CD (1.392)—not 80.0% and 52.6%. On Beauty, TRWH's best MAE (0.8421) is worse than ChatGPT few-shot (0.6977), so the claim of '10.8% improvement' and 'outperforms SOTA' is contradicted by the paper's own table. The abstract needs to specify the exact comparison baseline and recompute the percentages.
  3. [§3.2; §5.3] The paper does not state whether the heterogeneous graph is rebuilt per fold so that test interactions are excluded from edges. The graph is built from 'cleaned data' with rating/review/purchase edges between all users and items, and rating prediction is described only as the model 'outputs the predicted rating.' If test user–item edges remain in the graph during training/evaluation, the GNN has direct access to the target rating's edge, providing a second leakage path beyond the LLM profile issue. The evaluation protocol must be specified.
  4. [§5.3; Table 2] No error bars, confidence intervals, or multiple-seed results are reported, yet the NeurIPS checklist (item 7) answers 'Yes' to statistical significance. Many of the comparisons among TRWH variants are within ~1% (e.g., Fashion RMSE 1.0731 vs. 1.0604), so without variance estimates it is unclear whether the claimed advantages of LLMHet over W2VRHet, or of random-walk augmentation, are significant. This is essential for the paper's central empirical claims.
  5. [§4.2.1; Table 2] Most baseline methods are cross-domain recommenders (P2MF, APAR, PTUPCDR, CDRIB, UniCDR, NMCDR, RealHNS, MAN, PEMF-CD) evaluated on a single-domain rating task. Several report RMSE values above 7 on a 1–5 scale, far worse than simple MF (1.20 on Beauty), which suggests they are not adapted or tuned for this task. The resulting '80% reduction' claims are therefore obtained against non-competitive baselines. A fair SOTA comparison requires tuned single-domain baselines (e.g., MF, SVD++, NeuMF, LightGCN) on both datasets.
minor comments (5)
  1. [§3.1.2] 'Full details ... are provided in (create a figure)' is an unfinished placeholder; the prompt templates and examples must be included.
  2. [§2.2; §3.1.2] Typos: 'drown from RLMRec' should be 'drawn from'; 'we user an embedder' should be 'we use an embedder'.
  3. [§3.3] Edge names are inconsistent: the text says 'user-rated_by-item' and 'item-related_by-item' while §3.2 defines 'user-related_to-user' and 'item-related_to-item'.
  4. [Table 2] The two sub-tables are not cleanly separated; some entries are merged (e.g., '1.07310.9089'), making the results hard to read.
  5. [§5.3] The hyperparameter search is described but the model selection criterion is not stated; the paper should clarify whether reported numbers are from a validation split or the test fold.

Circularity Check

1 steps flagged

Evaluation may be self-referential: LLM profiles and graph edges are built from the full interaction history, so test-review text (and possibly the test rating edge) can enter the predictor; the paper never states per-fold regeneration.

specific steps
  1. other [Section 3.1.2, Eq. (3); Section 3.2 'Graph construction'; Section 5.3]
    "For each item v∈ Iu, we form a textual attribute cv = [α, θ, r^u_v], where r^u_v denotes the review written by user u for item v. The user prompt Q_u is then formulated as: Q_u = f_u({c_v | v∈I_u}). ... Full details of the prompt design—including S, Q, and f_u/f_v(·)—as well as illustrative examples, are provided in (create a figure). ... The graph is constructed from the cleaned data ... 'user-rates-item' ... edges link users to items based on the ratings they have provided. ... We employ 5-fold cross-validation to ensure robust evaluation."

    Eq. (3) puts the review text r^u_v of every interacted item into the user profile. The held-out rating to be predicted is exactly the rating for one such (u,v) interaction, and its review text is the strongest textual signal of that rating. The graph-construction passage builds rating/review edges from 'the cleaned data' with no per-fold exclusion, and §5.3 mentions only '5-fold cross-validation' with no statement that profiles or edges are regenerated from training-only interactions. As written, the test interaction's review text (and possibly its rating edge) is an input to the model that produces the 'prediction.' The deferred '(create a figure)' means the prompt details needed to disprove this reduction are absent. The evaluation is therefore self-referential unless an unstated per-fol

full rationale

This paper does not contain a mathematical derivation whose steps could be circular by construction; there are no self-citations, no imported uniqueness theorems, and the architecture is a combination of existing components (LLM profiling, HeteroGNN, random walks) with the relevant inspirations cited. The central claim is empirical: TRWH achieves large RMSE/MAE gains. The main circularity concern is therefore in the evaluation loop. The paper's own Eq. (3) constructs user profiles from the review text r^u_v of every item in I_u, and Section 3.2 constructs the graph from 'the cleaned data' including all rating/review edges. Section 5.3 reports '5-fold cross-validation' but never states that profiles are regenerated per fold or that test interactions are removed from I_u, α, θ, or r before profile generation and graph construction. If profiles and edges are built once from the full data, the predicted rating for a held-out interaction is produced from an input that already contains that interaction's review text and possibly the rating edge itself, so the reported RMSE/MAE could substantially reflect memorization rather than generalization. The placeholder '(create a figure)' in Section 3.1.2 means the prompt details needed to rule out this leakage are omitted. Separately, the abstract's headline reductions (80.0% RMSE and 52.6% MAE on Fashion; 25.7% and 10.8% on Beauty) are not directly derivable from Table 2, which is a reporting inconsistency rather than a circular-derivation issue. Weighing all of this, there is a genuine but partial self-referentiality risk in the empirical pipeline: the evaluation may reduce to reading the input if the unstated per-fold regeneration was not performed. Because the paper is otherwise self-contained and not built on self-citation, a moderate score of 4 is appropriate rather than a higher score that would require proof of actual label leakage.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

No new physical or conceptual entities are introduced. The load-bearing ingredients are the LLM profile prompts, the one-hop random-walk augmentation choice, and the split-unaware profile generation assumption, all of which are either fitted heuristics or unstated assumptions.

free parameters (6)
  • hidden_channels = 16, 32, 64, 128 (per model/dataset)
    Swept in Figure 4; the final choice 'depends on the specific model and dataset' (§5.3), meaning the best test-performing value is reported without a validation-based selection rule.
  • learning_rate = 1e-3
    Chosen from {1e-3, 1e-4, 1e-5} (§5.3) without a validation selection protocol.
  • epochs = 900 or 1000
    Selected from {300, 500, 800, 900, 1000, 1500, 2000} (§5.3); early stopping threshold 0.05 is ad hoc.
  • gnn_layers = 1
    Varied across {1, 2, 3}; best (1) reported.
  • word2vec_window_and_min_count = 5 / 1
    Set manually in §3.1.1.
  • random_walk_edges_per_node = 1
    One-hop walk selects exactly one second-order neighbor per node (§3.3); no sensitivity analysis to the number of sampled neighbors.
axioms (4)
  • domain assumption Amazon 2023 metadata and review text contain enough signal for LLM profiles to be semantically accurate and for the embedder to transfer it.
    Justifies the entire embedding creation pipeline (§3.1.2) and is not independently validated.
  • ad hoc to paper One-hop random walks create 'semantically proximal' second-order edges without noise propagation.
    Central to the augmentation claim (§3.3); not derived or validated beyond the two datasets.
  • domain assumption LLM profiles are generated from training-only interactions within each cross-validation fold.
    This is load-bearing for avoiding target leakage, but the paper never states that profiles are regenerated per fold (Section 3.1.2 and 5.3).
  • ad hoc to paper HeteroGNN message passing across 9 edge types preserves LLM semantics except when random-walk edges are added.
    The dilution finding is given as a post-hoc explanation (§4.2.2) rather than derived from first principles.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of TRWH: A Text-Driven Random Walk Heterogeneous GNN for Semantic-Aware Sparse Recommendation." pith.science (2026). https://pith.science/paper/CBN42SWW

@misc{pith2026260725471,
  author       = {Pith},
  title        = {Pith review of: TRWH: A Text-Driven Random Walk Heterogeneous GNN for Semantic-Aware Sparse Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CBN42SWW}},
  note         = {Machine review of arXiv:2607.25471}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Graph Neural Networks (GNNs) and Large Language Models (LLMs) have each advanced recommendation systems by modeling structural and semantic signals, respectively. However, integrating their complementary strengths remains challenging, particularly in sparse settings where maintaining semantic precision is critical. We propose TRWH (Text-driven Random Walk Heterogeneous Graph Neural Network), a novel framework that fuses LLM-generated textual profiles with heterogeneous graph structures through strategic random walk augmentation. TRWH consists of three core components: (1) Embedding Creation, which produces user and item representations using both Word2Vec and LLM-based profiling; (2) a Heterogeneous Graph Neural Network (HeteroGNN) that propagates information across multi-relational edges; and (3) Random Walk-based Path Construction, which enriches sparse graphs with second-order user-user and item-item links. Experiments on the Amazon-2023 Fashion (2M users, 825K items) and Beauty (631K users, 112K items) datasets demonstrate that TRWH achieves substantial performance gains over state-of-the-art methods, including 80.0% RMSE and 52.6% MAE reductions on Fashion, and 25.7% and 10.8% improvements on Beauty. Notably, while random walks improve performance with traditional embeddings, they can dilute the nuanced representations learned by LLMs, underscoring the importance of adaptive integration strategies.

Figures

Figures reproduced from arXiv: 2607.25471 by Chen Liu, He Ma.

Figure 1
Figure 1. Figure 1: Text-driven Random Walk Heterogeneous graph neural network(TRWH) main structure [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The method of applying LLM in our model 3.1.2 LLM-based approach To enhance the expressiveness of item and user representations, inspired by RLMRec [27], we leverage LLMs to generate semantic profiles for users and items. With the help of LLMs, we obtain 4 [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Random walk figure - it describes how we utilize random walk in our model [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of RMSE and MAE across different hidden channel sizes on the Beauty and [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

38 extracted references · 4 linked inside Pith

  1. [1]

    Enhancing cross-domain recommendations: Leveraging personality-based transfer learning with probabilistic matrix factorization.Expert Systems with Applications, 263:125667, 2025

    Somdeep Acharyya and Nargis Pervin. Enhancing cross-domain recommendations: Leveraging personality-based transfer learning with probabilistic matrix factorization.Expert Systems with Applications, 263:125667, 2025

    2025

  2. [2]

    Item2vec: neural item embeddings for collaborative filtering

    Oren Barkan and Noam Koenigstein. Item2vec: neural item embeddings for collaborative filtering. In2016 IEEE 26th International Workshop on Machine Learning for Signal Processing (MLSP), pages 1–6. IEEE, 2016

    2016

  3. [3]

    Towards universal cross-domain recommendation

    Jiangxia Cao, Shaoshuai Li, Bowen Yu, Xiaobo Guo, Tingwen Liu, and Bin Wang. Towards universal cross-domain recommendation. InProceedings of the sixteenth ACM international conference on web search and data mining, pages 78–86, 2023

    2023

  4. [4]

    Cross-domain recommen- dation to cold-start users via variational information bottleneck

    Jiangxia Cao, Jiawei Sheng, Xin Cong, Tingwen Liu, and Bin Wang. Cross-domain recommen- dation to cold-start users via variational information bottleneck. In2022 IEEE 38th International Conference on Data Engineering (ICDE), pages 2209–2223. IEEE, 2022

    2022

  5. [5]

    Wide & deep learning for recommender systems

    Heng-Tze Cheng, Levent Koc, Jeremiah Harmsen, Tal Shaked, Tushar Chandra, Hrishi Aradhye, Glen Anderson, Greg Corrado, Wei Chai, Mustafa Ispir, et al. Wide & deep learning for recommender systems. InProceedings of the 1st workshop on deep learning for recommender systems, pages 7–10, 2016

    2016

  6. [6]

    Rakcr: Reviews sentiment-aware based knowledge graph convolutional networks for personalized recommendation.Expert Systems with Applications, 248:123403, 2024

    Yachao Cui, Hongli Yu, Xiaoxu Guo, Han Cao, and Lei Wang. Rakcr: Reviews sentiment-aware based knowledge graph convolutional networks for personalized recommendation.Expert Systems with Applications, 248:123403, 2024

    2024

  7. [7]

    Homogeneous and heterogeneous relational graph for visible-infrared person re-identification

    Yujian Feng, Feng Chen, Jian Yu, Yimu Ji, Fei Wu, Shangdon Liu, and Xiao-Yuan Jing. Homogeneous and heterogeneous relational graph for visible-infrared person re-identification. Pattern Recognition, 158:110981, 2025

    2025

  8. [8]

    Chat-rec: Towards interactive and explainable llms-augmented recommender system

    Yunfan Gao, Tao Sheng, Youlin Xiang, Yun Xiong, Haofen Wang, and Jiawei Zhang. Chat-rec: Towards interactive and explainable llms-augmented recommender system. arxiv 2023.arXiv preprint arXiv:2303.14524, 2023

    Pith/arXiv arXiv 2023

  9. [9]

    Recommendation as language processing (rlp): A unified pretrain, personalized prompt & predict paradigm (p5)

    Shijie Geng, Shuchang Liu, Zuohui Fu, Yingqiang Ge, and Yongfeng Zhang. Recommendation as language processing (rlp): A unified pretrain, personalized prompt & predict paradigm (p5). InProceedings of the 16th ACM conference on recommender systems, pages 299–315, 2022. 12

    2022

  10. [10]

    node2vec: Scalable feature learning for networks

    Aditya Grover and Jure Leskovec. node2vec: Scalable feature learning for networks. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, pages 855–864, 2016

    2016

  11. [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. InProceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval, pages 639–648, 2020

    2020

  12. [12]

    Neural collaborative filtering

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

    2017

  13. [13]

    Towards universal sequence representation learning for recommender systems

    Yupeng Hou, Shanlei Mu, Wayne Xin Zhao, Yaliang Li, Bolin Ding, and Ji-Rong Wen. Towards universal sequence representation learning for recommender systems. InProceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 585–593, 2022

    2022

  14. [14]

    Tag-aware attentional graph neural networks for personalized tag recommendation

    Ruoran Huang, Chuanqi Han, and Li Cui. Tag-aware attentional graph neural networks for personalized tag recommendation. In2021 International joint conference on neural networks (IJCNN), pages 1–8. IEEE, 2021

    2021

  15. [15]

    Matrix factorization techniques for recom- mender systems.Computer, 42(8):30–37, 2009

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

    2009

  16. [16]

    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

  17. [17]

    Mixed attention network for cross-domain sequential recom- mendation

    Guanyu Lin, Chen Gao, Yu Zheng, Jianxin Chang, Yanan Niu, Yang Song, Kun Gai, Zhiheng Li, Depeng Jin, Yong Li, et al. Mixed attention network for cross-domain sequential recom- mendation. InProceedings of the 17th ACM international conference on web search and data mining, pages 405–413, 2024

    2024

  18. [18]

    How can recommender systems benefit from large language models: A survey.ACM Transactions on Information Systems, 43(2):1–47, 2025

    Jianghao Lin, Xinyi Dai, Yunjia Xi, Weiwen Liu, Bo Chen, Hao Zhang, Yong Liu, Chuhan Wu, Xiangyang Li, Chenxu Zhu, et al. How can recommender systems benefit from large language models: A survey.ACM Transactions on Information Systems, 43(2):1–47, 2025

    2025

  19. [19]

    Is chatgpt a good recommender? a preliminary study.arXiv preprint arXiv:2304.10149, 2023

    Junling Liu, Chao Liu, Peilin Zhou, Renjie Lv, Kang Zhou, and Yan Zhang. Is chatgpt a good recommender? a preliminary study.arXiv preprint arXiv:2304.10149, 2023

    Pith/arXiv arXiv 2023

  20. [20]

    Pre-train, prompt, and recommendation: A comprehensive survey of language modeling paradigm adaptations in recommender systems

    Peng Liu, Lemei Zhang, and Jon Atle Gulla. Pre-train, prompt, and recommendation: A comprehensive survey of language modeling paradigm adaptations in recommender systems. Transactions of the Association for Computational Linguistics, 11:1553–1571, 2023

    2023

  21. [21]

    Poi recommendation for random groups based on cooperative graph neural networks.Information Processing & Management, 61(3):103676, 2024

    Zhizhong Liu, Lingqiang Meng, Quan Z Sheng, Dianhui Chu, Jian Yu, and Xiaoyu Song. Poi recommendation for random groups based on cooperative graph neural networks.Information Processing & Management, 61(3):103676, 2024

    2024

  22. [22]

    Sequential recom- mendation with metric models based on frequent sequences.Data Mining and Knowledge Discovery, 35:1087–1133, 2021

    Corentin Lonjarret, Roch Auburtin, Céline Robardet, and Marc Plantevit. Sequential recom- mendation with metric models based on frequent sequences.Data Mining and Knowledge Discovery, 35:1087–1133, 2021

    2021

  23. [23]

    Exploring false hard negative sample in cross-domain recommendation

    Haokai Ma, Ruobing Xie, Lei Meng, Xin Chen, Xu Zhang, Leyu Lin, and Jie Zhou. Exploring false hard negative sample in cross-domain recommendation. InProceedings of the 17th ACM conference on recommender systems, pages 502–514, 2023

    2023

  24. [24]

    From theory to practice: The evolution and comparative analysis of homogeneous vs

    Maryam Khanian Najafabadi, Rei-An Chen, Javad Rezazadeh, Amin Beheshti, and Nasrin Shabani. From theory to practice: The evolution and comparative analysis of homogeneous vs. heterogeneous graph neural networks in recommender systems.Neurocomputing, page 129446, 2025. 13

    2025

  25. [25]

    Improving the accuracy of collaborative filtering recommendations using clustering and associa- tion rules mining on implicit data.Computers in Human Behavior, 67:113–128, 2017

    Maryam Khanian Najafabadi, Mohd Naz’ri Mahrin, Suriayati Chuprat, and Haslina Md Sarkan. Improving the accuracy of collaborative filtering recommendations using clustering and associa- tion rules mining on implicit data.Computers in Human Behavior, 67:113–128, 2017

    2017

  26. [26]

    Deepwalk: Online learning of social repre- sentations

    Bryan Perozzi, Rami Al-Rfou, and Steven Skiena. Deepwalk: Online learning of social repre- sentations. InProceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 701–710, 2014

    2014

  27. [27]

    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 Web Conference 2024, pages 3464–3475, 2024

    2024

  28. [28]

    One embedder, any task: Instruction-finetuned text embeddings.arXiv preprint arXiv:2212.09741, 2022

    Hongjin Su, Weijia Shi, Jungo Kasai, Yizhong Wang, Yushi Hu, Mari Ostendorf, Wen-tau Yih, Noah A Smith, Luke Zettlemoyer, and Tao Yu. One embedder, any task: Instruction-finetuned text embeddings.arXiv preprint arXiv:2212.09741, 2022

    Pith/arXiv arXiv 2022

  29. [29]

    Cross-domain recommendation with user personality.Knowledge-Based Systems, 213:106664, 2021

    Hanfei Wang, Yuan Zuo, Hong Li, and Junjie Wu. Cross-domain recommendation with user personality.Knowledge-Based Systems, 213:106664, 2021

    2021

  30. [30]

    Knowledge-aware graph neural networks with label attention mechanism for recommendation

    Hongwei Wang, Miao Zhao, Xing Xie, Wenjie Li, and Minlie Guo. Knowledge-aware graph neural networks with label attention mechanism for recommendation. InProceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCNLP), pages 3154–3164, 2019

    2019

  31. [31]

    Neural graph collabo- rative filtering

    Xiang Wang, Xiangnan He, Meng Wang, Fuli Feng, and Tat-Seng Chua. Neural graph collabo- rative filtering. InProceedings of the 42nd international ACM SIGIR conference on research and development in information retrieval, pages 165–174, 2019

    2019

  32. [32]

    Het- erogeneous graph attention network

    Xiao Wang, Houye Ji, Chuan Shi, Bai Wang, Yanfang Ye, Peng Cui, and Philip S Yu. Het- erogeneous graph attention network. InThe world wide web conference, pages 2022–2032, 2019

    2022

  33. [33]

    A survey on large language models for recommendation

    Likang Wu, Zhi Zheng, Zhaopeng Qiu, Hao Wang, Hongchao Gu, Tingjia Shen, Chuan Qin, Chen Zhu, Hengshu Zhu, Qi Liu, et al. A survey on large language models for recommendation. World Wide Web, 27(5):60, 2024

    2024

  34. [34]

    Neural node matching for multi-target cross domain recommendation

    Wujiang Xu, Shaoshuai Li, Mingming Ha, Xiaobo Guo, Qiongxu Ma, Xiaolei Liu, Linxun Chen, and Zhenfeng Zhu. Neural node matching for multi-target cross domain recommendation. In2023 IEEE 39th International Conference on Data Engineering (ICDE), pages 2154–2166. IEEE, 2023

    2023

  35. [35]

    Enabling the analysis of personality aspects in recommender systems.arXiv preprint arXiv:2001.04825, 2020

    Shahpar Yakhchi, Amin Beheshti, Seyed Mohssen Ghafari, and Mehmet Orgun. Enabling the analysis of personality aspects in recommender systems.arXiv preprint arXiv:2001.04825, 2020

    Pith/arXiv arXiv 2001

  36. [36]

    Recommender systems in the era of large language models (llms).IEEE Transactions on Knowledge and Data Engineering, 2024

    Zihuai Zhao, Wenqi Fan, Jiatong Li, Yunqing Liu, Xiaowei Mei, Yiqi Wang, Zhen Wen, Fei Wang, Xiangyu Zhao, Jiliang Tang, et al. Recommender systems in the era of large language models (llms).IEEE Transactions on Knowledge and Data Engineering, 2024

    2024

  37. [37]

    Limitations

    Yongchun Zhu, Zhenwei Tang, Yudan Liu, Fuzhen Zhuang, Ruobing Xie, Xu Zhang, Leyu Lin, and Qing He. Personalized transfer of user preferences for cross-domain recommendation. InProceedings of the fifteenth ACM international conference on web search and data mining, pages 1507–1515, 2022. 14 NeurIPS Paper Checklist 1.Claims Question: Do the main claims mad...

    2022

  38. [38]

    Guidelines: • The answer NA means that the paper does not involve crowdsourcing nor research with human subjects

    Institutional review board (IRB) approvals or equivalent for research with human subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or ...

    2025

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.