Representation Quantization for Collaborative Filtering Augmentation

Gaode Chen; Han Li; Jiang Yuezihan; Jian Liang; Jingchi Wang; Jun Zhang; Kaigui Bian; Kun Gai; Ruina Sun; Shicheng Wang

REVIEW 4 major objections 6 minor 52 references

Representation Quantization for Collaborative Filtering Augmentation

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 DQRec claims that quantizing ID-free user and item embeddings into SVD-decorrelated codebook IDs, then using those semantic patterns for feature and linkage augmentation, outperforms twelve baselines on three collaborative filtering dataset

desk verdict Genuinely different VQ scheme for CF augmentation with surprisingly large gains, but an ablation gap leaves the mechanism untested and the evaluation protocol inflates the headline numbers. read the letter →

arxiv 2508.11194 v1 pith:YFB6YRBX submitted 2025-08-15 cs.IR

Yunze Luo , Yinjie Jiang , Gaode Chen , Jingchi Wang , Shicheng Wang , Ruina Sun , Jiang Yuezihan , Jun Zhang

show 4 more authors

Jian Liang Han Li Kun Gai Kaigui Bian

This is my paper · ORCID

classification cs.IR

keywords recommendersystemscollaborativefilteringvectorquantizationsemanticIDsdatasparsitylinkageaugmentationsingularvaluedecompositionrepresentationlearning

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 claims that the chronic shortage of user-user and item-item links in collaborative filtering can be relieved by turning behavior into discrete codes. Its method, DQRec, first encodes users and items from features and interaction sequences alone (no IDs) into representation embeddings, then a new quantizer, DQ-VAE, linearly decorrelates those embeddings with an SVD and snaps the decorrelated coordinates into per-layer codebooks, producing a short semantic ID sequence for each user and item. Those IDs are treated as extracted behavior patterns and used twice: as extra user and item features, and as the basis for finding pattern-similar neighbors that create the missing homogeneous links. The paper reports that on MovieLens, Book, and Netflix this beats twelve baselines, with the largest relative gains in Recall (MovieLens Recall@10 0.8333 versus 0.6843 for the best baseline), and that the gain shrinks on datasets with short interaction sequences such as Book. A sympathetic reader would care because the method captures similarity that overlapping-interaction and shared-attribute heuristics miss, and because each dimension of the semantic ID is meant to be interpretable and independent.

What carries the argument

DQ-VAE is the carrying mechanism. It computes the SVD $\tilde{Z} = V\Sigma W^\top$ of the centralized embedding matrix, so the transformed coordinates $Y = \tilde{Z}W$ have diagonal covariance and per-dimension variance proportional to $\sigma_n^2$ (Eqs. 5-6); $W$ is then permuted into $L$ column-blocks $W_l$ whose squared-singular-value sums are balanced (Eq. 7). Each block acts as a fixed linear encoder for one layer, $x_l = (z - \bar{z})W_l$, and nearest-codebook lookup assigns the codeword $c_l$ (Eq. 9); the $L$ codewords concatenate into the semantic ID $s = \{c_1, \ldots, c_L\}$. Because the encoder comes from SVD, training needs no stop-gradient — only a reconstruction loss and a code

What would settle it

Run the full DQRec pipeline but replace the codebook lookup in Eq. (9) with random codeword assignment (same codebooks, same neighbor selection, same augmentation); if Recall@10 on MovieLens stays near the reported 0.8333, the semantic IDs themselves are not the source of the gain. A second check: after training, permute the semantic IDs across users and items and re-evaluate — if scores barely drop, the IDs carry no pattern information and the augmentation machinery alone explains the result.

Watch

Extended reading notes

Core claim

DQRec's central claim is that behavior patterns, such as a user's multi-aspect interests, can be extracted jointly from interaction sequences and attributes by quantizing a pre-trained representation, and that these patterns, encoded as semantic ID sequences, can be fed back into collaborative filtering as both enriched features and the basis for new homogeneous linkages. The decisive step is DQ-VAE: the centered embedding matrix is decomposed by SVD, each embedding is mapped through column blocks of the right singular matrix chosen so that every layer carries comparable information (balanced sums of squared singular values), and nearest-codebook lookup assigns one codeword per layer, yieldi

Load-bearing premise

The entire pipeline rests on the assumption that the pretrained, ID-free embeddings carry genuine behavior patterns in directions that a linear SVD decorrelation can separate and a codebook can snap to; if those directions do not align with meaningful aspects of user or item behavior, the semantic IDs become arbitrary labels and the augmentation adds no signal beyond the original embeddings.

Editorial extensions

If this is right

Homogeneous-linkage augmentation no longer has to wait for overlapping interacted neighbors or shared coarse attributes: quantized codes supply pattern-level similarity, and the paper's case study shows DQRec both creates links where sequences barely overlap and refuses links where sequences overlap but interests differ.
Across all three datasets the relative Recall gains exceed the NDCG gains (e.g., Netflix Recall@10 +27.50% versus NDCG@10 +16.56%), which the paper reads as DQRec strengthening retrieval rather than re-ranking precision.
The pretrained DQ-VAE is frozen before recommender training, keeping online cost at one codebook lookup plus neighbor aggregation: total inference complexity is $O(Ld + |S|d + Kd + d^2)$ with average pooling, as stated in the paper.
The paper's own results bound the claim: on the Book dataset, where interaction sequences are short, the improvement is smaller because there is less sequence evidence from which to form patterns.
Ablations removing either user-side or item-side feature augmentation, or latent/user/item linkage augmentation, all lower Recall@5 and NDCG@5, so the reported gain is not carried by any single augmentation pathway.

Reading between the lines

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

The paper leaves implicit that because the pretrained embeddings are built from features and neighbors rather than IDs, DQRec's semantic IDs should also exist for cold-start users and items; a warm-versus-cold slice of the same datasets would test whether the pattern codes, not the ID embeddings, are what transfers.
The latent-pattern step swaps a single codeword for its runner-up in one dimension; an untested extension is flipping codewords in several dimensions at once, which would show whether the discovered neighbor sets are an artifact of the one-at-a-time search.
The SVD and codebooks are estimated once over the whole population and never re-fit; whether the gain survives distribution drift (re-estimating W and the codebooks on rolling windows) is an open, testable question.
Because the semantic IDs are discrete, decoupled, and interpretable per dimension, a natural next step the paper does not take is to feed the ID sequences as tokens into a generative or LLM-based recommender, extending the feature-augmentation role into a language-model input.

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.

Desk Editor's Note

read the letter

Short version: this is a plausible two-stage recommender method with a real novelty—SVD-based decorrelation of pretrained embeddings to build parallel, independent codebooks for users and items—and the reported gains are large. But the bigger the gain, the more I want to know what is causing it, and the paper doesn't tell us. The ablations delete whole feature or linkage blocks; they never swap the SVD+quantization for an ordinary VQ-VAE or even random codebook assignment, so the gains could come from the generic augmentation machinery (neighbor pooling and code-lookup features) rather than the 'pattern semantics' being sold. The stress-test note lands squarely: this missing control is load-bearing. If any quantizer gives the same lift, the novelty claim is mostly about the pipeline, not the decomposition.

What's genuinely good: applying VQ to both users and items rather than just items, and forcing dimensional decorrelation via SVD, is a clean idea. The interpretability story—each position in the ID corresponds to a distinct aspect of behavior—is convincing. The case study in the paper shows exactly the kind of detection they want. The method is clearly described, and the complexity analysis, though brief, shows they thought about inference.

The soft spots are mostly empirical. Evaluation uses only in-batch negative samples, which inflates absolute metrics and makes cross-paper comparison meaningless; internally it's consistent, but then you need error bars or multiple seeds—none are reported. Baseline tuning is under-specified: saying parameters 'remain the same' without saying same as what is not enough. Training cost is never discussed: computing top-K nearest users for every training sample is O(N) per step, and with 55k users that is heavy; they might use approximate nearest neighbors, but the paper is silent. Duplicated references ([19]=[18], [45]=[46]) are sloppy but minor. I'm less worried about the circularity charge in the reader's note—IDs from a pretrained model on the training set are just features; test labels are held out. The missing control is the bigger deal.

Who's this for: people building augmentation pipelines for CF will get ideas; the SVD-to-codebook trick is novel and easy to lift. But the empirical claims need a firmer base before I'd build on them. This deserves a serious referee, not a desk reject. The idea is worth engaging with. I'd push hard for an ablation that holds the augmentation fixed and randomizes the codebooks, and for a full-ranking evaluation on at least one dataset.

Referee Report

4 major / 6 minor

Summary. The paper proposes DQRec, a two-stage collaborative filtering method. In the first stage, a dual-tower model is pre-trained on user and item features plus neighbor features (without user/item IDs). The resulting embeddings are then fed into a proposed DQ-VAE, which uses SVD to decorrelate embedding dimensions (Eqs. 3–7), splits the transformation into L sub-matrices, quantizes each part through a separate codebook (Eqs. 8–9), and reconstructs a quantized embedding (Eq. 10). The resulting semantic ID sequences are used in the second stage for feature augmentation (Eq. 14) and for homogeneous linkage augmentation: top-K similar users/items are selected by distance in quantized space (Eq. 15), with additional 'latent' neighbors obtained by replacing one codeword with the second-best codebook entry (Eqs. 16–19). Experiments compare DQRec with 12 baselines on MovieLens, Books, and Netflix, reporting Recall@K and NDCG@K, plus ablations and parameter sensitivity studies. The paper reports large gains over all baselines, e.g., MovieLens Recall@10 0.8333 vs. 0.6843 for VQRec.

Significance. The SVD-based decomposition is an interesting and potentially elegant alternative to hierarchical RQ-VAE: it gives closed-form, parallel codebook layers that are intended to be decorrelated and interpretable, and the paper provides a clear derivation of the encoder/decoder construction. The case study in Section 6.5 gives qualitative evidence that the resulting semantic IDs capture category-level interest patterns better than two-hop overlap. If the reported gains are real and attributable to the DQ-VAE mechanism, the method would be a useful contribution to collaborative-filtering augmentation. However, the empirical support is currently weakened by the evaluation protocol and by the absence of a control that isolates the DQ-VAE's semantic decomposition from generic augmentation. The paper does not supply machine-checked proofs or code, so the reproducibility rests on the clarity of the experimental description; several implementation details (e.g., how Eq. (7) is solved) are underspecified.

major comments (4)

[Section 6.1.3] The evaluation protocol computes Recall@K and NDCG@K using in-batch negative sampling: 'all the non-interacted items in the current batch as negative samples.' This is not a full-ranking protocol, and it can inflate absolute metric values and change relative model rankings depending on batch composition. The claim of 'superior performance' over 12 baselines needs a consistent full-ranking evaluation or a fixed sampled-negative protocol, with results averaged over multiple seeds and significance tests. Without this, the magnitude of the improvements in Table 1 is not trustworthy.
[Section 6.1.1] The datasets are split into training, validation, and test sets 'in a ratio of 8:1:1.' Because the method explicitly uses recent interaction sequences to build user/item representations, a random split of interactions can leak future interactions into the training sequences of test samples. If the split is random rather than temporal, the reported gains may overstate the method's ability to model sequential behavior. The authors should specify whether the split is by timestamp or by user, and if it is random, re-run with a temporal split or justify why leakage is not present.
[Table 3 / Sections 4.2.2–4.3.2] The ablation study removes whole feature-augmentation or relation-augmentation blocks, but it never tests whether the specific DQ-VAE mechanism matters. A reader cannot tell whether the gains come from the SVD-decorrelated, dimension-decoupled quantization claimed as the novelty, or simply from adding any codebook-lookup features and nearest-neighbor aggregation on top of the pre-trained embeddings. A necessary control is to replace DQ-VAE with (a) a standard VQ-VAE/RQ-VAE using the same embedding dimension and codebook sizes, (b) randomly assigned codebooks, and (c) continuous nearest-neighbor retrieval on the raw pre-trained embeddings, while keeping the downstream augmentation components unchanged. Without these controls, the causal story behind Eqs. (3)–(19) is not established.
[Section 4.2.1] The semantic IDs are generated from a pre-trained model and a DQ-VAE trained on the same data as the final recommendation model. This does not make the held-out test evaluation circular, but it weakens the claim that the method extracts 'behavior characteristics jointly from interaction sequences and attributes' as an independent source of signal. The pre-trained embeddings already encode the collaborative information from the training interactions, so the semantic IDs may simply be a re-encoding of the same signal. A concrete test would be to train the pre-trained model on a split disjoint from the DQ-VAE and downstream training, or to use only side information (no interaction sequences) for the pretrained embeddings, and then measure whether the augmentation still adds value.

minor comments (6)

[Section 5] The space complexity is stated as O((|U|+|I|)d + Ld + d^2), but the codebook contains L layers each with J embeddings of dimension d/L, giving O(Jd), not O(Ld). Additionally, the semantic feature embedding matrices in Eq. (14) appear to add a term that is not counted. Please correct the complexity expression.
[Eq. (7)] The balanced-column partitioning problem is defined as an arg min over column permutations, but the algorithm used to solve it is not described. Since the exact partition affects all downstream semantic IDs, please specify whether the assignment is exact, greedy, or solved by a heuristic.
[Section 6.1.4] The statement 'The parameters of the baselines remain the same' is too vague. For a fair comparison, the authors should state whether baseline hyperparameters were tuned on the validation set and report the chosen values or cite the original papers' defaults.
[Figure 3] The y-axis is labeled 'Performance' and is normalized to 80–105%, but the absolute metric being plotted and the reference point for 100% are not defined. Please report the actual metric values or clearly state the normalization.
[References] Reference [19] and reference [46] are duplicated (both are 'Large Language Models meet Collaborative Filtering' / 'Adapting Large Language Models...'). Reference [1] appears to have an incorrect author name ('Agfarap' instead of 'Agarap') and duplicates reference [10].
[Eq. (20)] The user input embedding contains ID embeddings, semantic ID embeddings, aggregated neighbor ID embeddings, and aggregated similar-user ID embeddings, but no explicit attribute embeddings. Since the pre-trained model uses attributes, please clarify how attributes are carried into the final dual-tower model, or whether the semantic ID embeddings are the only attribute-derived signal.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: semantic IDs are unsupervised quantizations of pre-trained training-set embeddings, and evaluation is on held-out interactions.

full rationale

DQRec's derivation chain is self-contained and does not reduce to its own inputs. The pre-trained dual-tower model (Sec. 4.2.1) is trained on user/item features plus neighbor features, without IDs, producing representation embeddings z. DQ-VAE (Secs. 4.2.2-4.2.3) computes an SVD-based linear transformation W from the training-set embedding matrix (Eqs. 2-7) and optimizes codebooks with reconstruction and commitment losses (Eqs. 11-13). Semantic IDs are then obtained by nearest-neighbor lookup (Eq. 9) and quantized representations by Eq. 10. These IDs are used to build augmented features (Eq. 14) and top-K neighbor sets (Eqs. 15-19) for the final dual-tower recommender (Eqs. 20-23). At no point is the prediction y_hat used to construct the semantic IDs, nor is any fitted parameter renamed as a prediction; the final model is trained with BPR on the training split and evaluated on a held-out split. The SVD/codebook training is unsupervised with respect to the recommendation labels, so the quantized patterns are a compression of the pre-trained representations rather than a circular restatement of the test labels. The only self-citations are background references (e.g., [4], [24]) and are not load-bearing. The absence of a control replacing DQ-VAE with ordinary VQ-VAE is an ablation/experimental-design limitation, not circularity, and does not change the independence of the derivation chain.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The main free parameters are the codebook layers, codebook size, neighbor counts, and commitment weight, all tuned on validation. The SVD step is standard math. The key domain assumption is that pre-trained embeddings encode semantically meaningful patterns that survive quantization. The latent-interest exploration is an ad hoc heuristic.

free parameters (6)

L (number of codebook layers) = 4
Tuned on validation; Fig. 3c shows performance degrades for L=2 or L=5. The paper states L=4 without discussing sensitivity on other datasets.
J (codebook size) = 128
Tuned on validation; Fig. 3b shows J=32/64 lower performance and J=256 also lower; J=128 is chosen.
K (number of similar neighbors) = 30
Tuned; Fig. 3a shows K=10 or K=40 worse; K=30 chosen.
K' (latent neighbors per dimension) = 2
Set by hand, no sensitivity analysis provided.
beta (commitment weight) = 0.25
Hyperparameter in Eq. 13; no analysis provided.
embedding dimension d = 64
Common setting used across all embeddings, no analysis.

assumptions (4)

standard math SVD orthogonality and decorrelation properties
Eqs. 3-7 rely on the fact that W from SVD is orthogonal and diagonalizes the covariance matrix; this is textbook linear algebra.
domain assumption Pre-trained embeddings capture behavior patterns
Section 4.2.1 assumes the dual-tower model produces embeddings that encode useful behavior patterns without using IDs; this is not independently validated.
ad hoc to paper Quantization preserves pattern semantics
The paper assumes nearest-codebook lookup (Eq. 9) yields semantic IDs that reflect true user/item aspects; no token-level validation beyond one qualitative case.
ad hoc to paper Latent pattern exploration via second-best codeword
Eq. 16 assumes the second-nearest codeword corresponds to an alternative genuine interest, not noise; no theoretical or empirical justification is given.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Representation Quantization for Collaborative Filtering Augmentation." pith.science (2026). https://pith.science/paper/YFB6YRBX

@misc{pith2026250811194,
  author       = {Pith},
  title        = {Pith review of: Representation Quantization for Collaborative Filtering Augmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YFB6YRBX}},
  note         = {Machine review of arXiv:2508.11194}
}

read the original abstract

As the core algorithm in recommendation systems, collaborative filtering (CF) algorithms inevitably face the problem of data sparsity. Since CF captures similar users and items for recommendations, it is effective to augment the lacking user-user and item-item homogeneous linkages. However, existing methods are typically limited to connecting through overlapping interacted neighbors or through similar attributes and contents. These approaches are constrained by coarse-grained, sparse attributes and fail to effectively extract behavioral characteristics jointly from interaction sequences and attributes. To address these challenges, we propose a novel two-stage collaborative recommendation algorithm, DQRec: Decomposition-based Quantized Variational AutoEncoder (DQ-VAE) for Recommendation. DQRec augments features and homogeneous linkages by extracting the behavior characteristics jointly from interaction sequences and attributes, namely patterns, such as user multi-aspect interests. Inspired by vector quantization (VQ) technology, we propose a new VQ algorithm, DQ-VAE, which decomposes the pre-trained representation embeddings into distinct dimensions, and quantize them to generates semantic IDs. We utilize the generated semantic IDs as the extracted patterns mentioned above. By integrating these semantic ID patterns into the recommendation process through feature and linkage augmentation, the system enriches both latent and explicit user and item features, identifies pattern-similar neighbors, and thereby improves the efficiency of information diffusion. Experimental comparisons with baselines across multiple datasets demonstrate the superior performance of the proposed DQRec method.

Figures

Figures reproduced from arXiv: 2508.11194 by the authors.

**Figure 2.** The detailed process of linkage augmentation and the collaborative filtering model. (1) Linkages are augmented through [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 3.** Performance of DQRec on MovieLens dataset [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Categories of items in interaction sequence of users, [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

52 extracted references · 13 canonical work pages

[19]

Sein Kim, Hongseok Kang, Seungyoon Choi, Donghyun Kim, Minchul Yang, and Chanyoung Park. 2024. Large Language Models meet Collaborative Filtering: An Efficient All-round LLM-based Recommender System. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (Barcelona, Spain) (KDD ’24). Association for Computing Machinery, Ne...

arXiv 2024
[46]

Bowen Zheng, Yupeng Hou, Hongyu Lu, Yu Chen, Wayne Xin Zhao, Ming Chen, and Ji-Rong Wen. 2024. Adapting Large Language Models by Integrating Collaborative Semantics for Recommendation. In 2024 IEEE 40th International Conference on Data Engineering (ICDE) . 1435–1448. https://doi.org/10.1109/ ICDE60146.2024.00118

arXiv 2024
[1]

Abien Fred Agfarap. 2019. Deep Learning using Rectified Linear Units (ReLU). arXiv:1803.08375 [cs.NE]

arXiv 2019
[2]

Jan Van Balen and Mark Levy. 2019. PQ-VAE: Efficient Recommendation Using Quantized Embeddings. In ACM Conference on Recommender Systems . https: //api.semanticscholar.org/CorpusID:201672459

work page 2019
[3]

Yi Cao, Sihao Hu, Yu Gong, Zhao Li, Yazheng Yang, Qingwen Liu, and Shoul- ing Ji. 2022. GIFT: Graph-guIded Feature Transfer for Cold-Start Video Click- Through Rate Prediction. In Proceedings of the 31st ACM International Con- ference on Information & Knowledge Management (Atlanta, GA, USA) (CIKM ’22). Association for Computing Machinery, New York, NY, US...

work page arXiv 2022
[4]

Gaode Chen, Ruina Sun, Yuezihan Jiang, Jiangxia Cao, Qi Zhang, Jingjian Lin, Han Li, Kun Gai, and Xinghua Zhang. 2024. A Multi-modal Modeling Framework for Cold-start Short-video Recommendation. In Proceedings of the 18th ACM Conference on Recommender Systems . 391–400

work page 2024
[5]

Yongjun Chen, Zhiwei Liu, Jia Li, Julian McAuley, and Caiming Xiong. 2022. Intent Contrastive Learning for Sequential Recommendation. In Proceedings of the ACM Web Conference 2022 (Virtual Event, Lyon, France) (WWW ’22) . Association for Computing Machinery, New York, NY, USA, 2172–2182. https: //doi.org/10.1145/3485447.3512090

work page arXiv 2022
[6]

Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. arXiv:1810.04805 [cs.CL] https://arxiv.org/abs/1810.04805

arXiv 2019

Show all 52 references

[7]

Maxwell Harper and Joseph A

F. Maxwell Harper and Joseph A. Konstan. 2015. The MovieLens Datasets: History and Context. ACM Trans. Interact. Intell. Syst. 5, 4, Article 19 (dec 2015), 19 pages. https://doi.org/10.1145/2827872

2015 doi
[8]

Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, YongDong Zhang, and Meng Wang. 2020. LightGCN: Simplifying and Powering Graph Convolution Network for Recommendation. In Proceedings of the 43rd International ACM SIGIR Confer- ence on Research and Development in Information Retrieva...

2020
[9]

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(Perth, Australia) (WWW ’17). International World Wide Web Conferences Steering Committee, R...

2017
[10]

Dan Hendrycks and Kevin Gimpel. 2023. Gaussian Error Linear Units (GELUs). arXiv:1606.08415 [cs.LG] https://arxiv.org/abs/1606.08415

2023 arXiv
[11]

Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk
[12]

Yupeng Hou, Zhankui He, Julian McAuley, and Wayne Xin Zhao. 2023. Learning vector-quantized item representation for transferable sequential recommenders. In Proceedings of the ACM Web Conference 2023 . 1162–1171

2023
[13]

Yupeng Hou, Jiacheng Li, Zhankui He, An Yan, Xiusi Chen, and Julian McAuley
[14]

Zheng Hu, Zhe Li, Ziyun Jiao, Satoshi Nakagawa, Jiawen Deng, Shimin Cai, Tao Zhou, and Fuji Ren. 2024. Bridging the User-side Knowledge Gap in Knowledge- aware Recommendations with Large Language Models. arXiv:2412.13544 [cs.IR] https://arxiv.org/abs/2412.13544

2024 arXiv
[15]

Po-Sen Huang, Xiaodong He, Jianfeng Gao, Li Deng, Alex Acero, and Larry Heck. 2013. Learning deep structured semantic models for web search using Representation Quantization for Collaborative Filtering Augmentation Conference acronym ’XX, June 03–05, 2018, Woodstock, NY clickt...

2013
[16]

Wei Jiang, Xinyi Gao, Guandong Xu, Tong Chen, and Hongzhi Yin. 2024. Chal- lenging Low Homophily in Social Recommendation. arXiv:2401.14606 [cs.IR] https://arxiv.org/abs/2401.14606

2024 arXiv
[17]

Wang-Cheng Kang and Julian McAuley. 2018. Self-Attentive Sequential Recom- mendation. arXiv:1808.09781 [cs.IR] https://arxiv.org/abs/1808.09781

2018 arXiv
[20]

Kingma and Jimmy Ba

Diederik P. Kingma and Jimmy Ba. 2014. Adam: A Method for Stochastic Opti- mization. https://doi.org/10.48550/ARXIV.1412.6980

2014 doi arXiv
[21]

Klema and A

V. Klema and A. Laub. 1980. The singular value decomposition: Its computation and some applications. IEEE Trans. Automat. Control 25, 2 (1980), 164–176. https://doi.org/10.1109/TAC.1980.1102314

1980
[22]

Yehuda Koren, Robert Bell, and Chris Volinsky. 2009. Matrix Factorization Techniques for Recommender Systems. Computer 42, 8 (2009), 30–37. https: //doi.org/10.1109/MC.2009.263

2009 doi
[23]

Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. 2022. Autoregressive Image Generation using Residual Quantization. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) . 11513–11522. https://doi.org/10.1109/CVPR52688.2022.01123

2022
[24]

Yunze Luo, Yuezihan Jiang, Yinjie Jiang, Gaode Chen, Jingchi Wang, Kaigui Bian, Peiyi Li, and Qi Zhang. 2025. Online Item Cold-Start Recommendation with Popularity-Aware Meta-Learning. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.1 ...

2025
[25]

Andrzej Maćkiewicz and Waldemar Ratajczak. 1993. Principal components analysis (PCA). Computers & Geosciences 19, 3 (1993), 303–342

1993
[26]

Ruihong Qiu, Zi Huang, Hongzhi Yin, and Zijian Wang. 2022. Contrastive Learn- ing for Representation Degeneration Problem in Sequential Recommendation. In Proceedings of the Fifteenth ACM International Conference on Web Search and Data Mining (Virtual Event, AZ, USA) (WSDM ’22...

2022
[27]

Shashank Rajput, Nikhil Mehta, Anima Singh, Raghunandan Hulikal Keshavan, Trung Vu, Lukasz Heldt, Lichan Hong, Yi Tay, Vinh Tran, Jonah Samost, et al
[28]

Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. 2019. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems 32 (2019)

2019
[29]

Steffen Rendle, Christoph Freudenthaler, Zeno Gantner, and Lars Schmidt-Thieme
[30]

Anima Singh, Trung Vu, Nikhil Mehta, Raghunandan Keshavan, Maheswaran Sathiamoorthy, Yilin Zheng, Lichan Hong, Lukasz Heldt, Li Wei, Devansh Tandon, Ed Chi, and Xinyang Yi. 2024. Better Generalization with Semantic IDs: A Case Study in Ranking for Recommendations. In Proceedin...

2024
[31]

Erfani, and Junhao Gan

Yixin Su, Rui Zhang, Sarah M. Erfani, and Junhao Gan. 2021. Neural Graph Matching based Collaborative Filtering. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval (Virtual Event, Canada) (SIGIR ’21). Association ...

2021
[32]

Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang
[33]

Jiaxi Tang and Ke Wang. 2018. Personalized Top-N Sequential Recommendation via Convolutional Sequence Embedding. In Proceedings of the Eleventh ACM International Conference on Web Search and Data Mining (Marina Del Rey, CA, USA) (WSDM ’18). Association for Computing Machinery,...

2018
[34]

Aaron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. 2018. Neural Discrete Representation Learning. arXiv:1711.00937 [cs.LG] https://arxiv.org/ abs/1711.00937

2018 arXiv
[35]

Gomez, Łukasz Kaiser, and Illia Polosukhin

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Proceedings of the 31st International Conference on Neural Information Processing Systems (Long Beach, California, ...

2017
[36]

Xiang Wang, Xiangnan He, Meng Wang, Fuli Feng, and Tat-Seng Chua. 2019. Neural Graph Collaborative Filtering. In Proceedings of the 42nd International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR ’19). ACM, 165–174. https://doi.org/10.1145/3...

2019
[37]

Yining Wang, Liwei Wang, Yuanzhi Li, Di He, Tie-Yan Liu, and Wei Chen. 2013. A Theoretical Analysis of NDCG Type Ranking Measures. arXiv:1304.6480 [cs.LG]

2013 arXiv
[38]

Chunyu Wei, Jian Liang, Di Liu, Zehui Dai, Mang Li, and Fei Wang. 2023. Meta Graph Learning for Long-tail Recommendation. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (Long Beach, CA, USA) (KDD ’23). Association for Computing Machiner...

2023
[39]

Shu Wu, Yuyuan Tang, Yanqiao Zhu, Liang Wang, Xing Xie, and Tieniu Tan. 2019. Session-Based Recommendation with Graph Neural Networks. Proceedings of the AAAI Conference on Artificial Intelligence 33, 01 (Jul. 2019), 346–353. https: //doi.org/10.1609/aaai.v33i01.3301346

2019 doi
[40]

Lianghao Xia, Chao Huang, Jiao Shi, and Yong Xu. 2023. Graph-less Collaborative Filtering. In Proceedings of the ACM Web Conference 2023 (Austin, TX, USA) (WWW ’23). Association for Computing Machinery, New York, NY, USA, 17–27. https://doi.org/10.1145/3543507.3583196

2023
[41]

Lianghao Xia, Chao Huang, Yong Xu, Jiashu Zhao, Dawei Yin, and Jimmy Huang
[42]

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) . 1259–1273. https://doi.org/10.1109/ICDE53745.2022.00099

2022
[43]

Yaowen Ye, Lianghao Xia, and Chao Huang. 2023. Graph Masked Autoencoder for Sequential Recommendation. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval (Taipei, Taiwan) (SIGIR ’23). Association for Computing Mac...

2023
[44]

Jiani Zhang, Xingjian Shi, Shenglin Zhao, and Irwin King. 2019. STAR-GCN: stacked and reconstructed graph convolutional networks for recommender sys- tems. In Proceedings of the 28th International Joint Conference on Artificial Intelli- gence (Macao, China) (IJCAI’19). AAAI Pr...

2019
[47]

Lei Zheng, Chaozhuo Li, Chun-Ta Lu, Jiawei Zhang, and Philip S. Yu. 2019. Deep Distribution Network: Addressing the Data Sparsity Issue for Top-N Recommendation. In Proceedings of the 42nd International ACM SIGIR Confer- ence on Research and Development in Information Retrieva...

2019
[48]

Hongmin Zhu, Fuli Feng, Xiangnan He, Xiang Wang, Yan Li, Kai Zheng, and Yongdong Zhang. 2020. Bilinear graph neural network with neighbor interactions. arXiv preprint arXiv:2002.03575 (2020)

2020 arXiv
[2009]

In Proceedings of the Twenty-Fifth Conference on Uncertainty in Artificial Intelligence (Montreal, Quebec, Canada) (UAI ’09)

BPR: Bayesian personalized ranking from implicit feedback. In Proceedings of the Twenty-Fifth Conference on Uncertainty in Artificial Intelligence (Montreal, Quebec, Canada) (UAI ’09). AUAI Press, Arlington, Virginia, USA, 452–461
[2016]

arXiv:1511.06939 [cs.LG] https://arxiv.org/abs/1511.06939

Session-based Recommendations with Recurrent Neural Networks. arXiv:1511.06939 [cs.LG] https://arxiv.org/abs/1511.06939

arXiv
[2019]

In Proceedings of the 28th ACM International Conference on Information and Knowledge Management (Beijing, China) (CIKM ’19)

BERT4Rec: Sequential Recommendation with Bidirectional Encoder Rep- resentations from Transformer. In Proceedings of the 28th ACM International Conference on Information and Knowledge Management (Beijing, China) (CIKM ’19). Association for Computing Machinery, New York, NY, US...
[2022]

In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval (Madrid, Spain) (SIGIR ’22)

Hypergraph Contrastive Collaborative Filtering. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval (Madrid, Spain) (SIGIR ’22). Association for Computing Machinery, New York, NY, USA, 70–79. https://doi.org/10.114...
[2023]

Advances in Neural Information Processing Systems 36 (2023), 10299–10315

Recommender systems with generative retrieval. Advances in Neural Information Processing Systems 36 (2023), 10299–10315

2023
[2024]

arXiv preprint arXiv:2403.03952 (2024)

Bridging Language and Items for Retrieval and Recommendation. arXiv preprint arXiv:2403.03952 (2024)

2024 arXiv

Pith tools

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

[1] [19]

Sein Kim, Hongseok Kang, Seungyoon Choi, Donghyun Kim, Minchul Yang, and Chanyoung Park. 2024. Large Language Models meet Collaborative Filtering: An Efficient All-round LLM-based Recommender System. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (Barcelona, Spain) (KDD ’24). Association for Computing Machinery, Ne...

arXiv 2024

[2] [46]

Bowen Zheng, Yupeng Hou, Hongyu Lu, Yu Chen, Wayne Xin Zhao, Ming Chen, and Ji-Rong Wen. 2024. Adapting Large Language Models by Integrating Collaborative Semantics for Recommendation. In 2024 IEEE 40th International Conference on Data Engineering (ICDE) . 1435–1448. https://doi.org/10.1109/ ICDE60146.2024.00118

arXiv 2024

[3] [1]

Abien Fred Agfarap. 2019. Deep Learning using Rectified Linear Units (ReLU). arXiv:1803.08375 [cs.NE]

arXiv 2019

[4] [2]

Jan Van Balen and Mark Levy. 2019. PQ-VAE: Efficient Recommendation Using Quantized Embeddings. In ACM Conference on Recommender Systems . https: //api.semanticscholar.org/CorpusID:201672459

work page 2019

[5] [3]

Yi Cao, Sihao Hu, Yu Gong, Zhao Li, Yazheng Yang, Qingwen Liu, and Shoul- ing Ji. 2022. GIFT: Graph-guIded Feature Transfer for Cold-Start Video Click- Through Rate Prediction. In Proceedings of the 31st ACM International Con- ference on Information & Knowledge Management (Atlanta, GA, USA) (CIKM ’22). Association for Computing Machinery, New York, NY, US...

work page arXiv 2022

[6] [4]

Gaode Chen, Ruina Sun, Yuezihan Jiang, Jiangxia Cao, Qi Zhang, Jingjian Lin, Han Li, Kun Gai, and Xinghua Zhang. 2024. A Multi-modal Modeling Framework for Cold-start Short-video Recommendation. In Proceedings of the 18th ACM Conference on Recommender Systems . 391–400

work page 2024

[7] [5]

Yongjun Chen, Zhiwei Liu, Jia Li, Julian McAuley, and Caiming Xiong. 2022. Intent Contrastive Learning for Sequential Recommendation. In Proceedings of the ACM Web Conference 2022 (Virtual Event, Lyon, France) (WWW ’22) . Association for Computing Machinery, New York, NY, USA, 2172–2182. https: //doi.org/10.1145/3485447.3512090

work page arXiv 2022

[8] [6]

Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. arXiv:1810.04805 [cs.CL] https://arxiv.org/abs/1810.04805

arXiv 2019