Pith. sign in

REVIEW 4 major objections 4 minor 48 references

Soft-token fusion lifts LLM recommender retrieval on three benchmarks.

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 ↗

T0 review · deepseek-v4-flash

2026-08-04 01:39 UTC pith:6VG6X7P6

load-bearing objection The interaction-vs-concatenation ablation holds up, but the abstract overclaims: the embedding stream is confounded with prompt richness and a 4B auxiliary encoder. the 4 major comments →

arxiv 2607.10016 v3 pith:6VG6X7P6 submitted 2026-07-10 cs.IR cs.LG

Tokenizing Numerical and Embedding Features for LLM RecSys

Zhe Xu , Ankit Peshin , Chiyu Zhang , Feng Qi , Johnson Lui , Anil Ramakrishna , Justin Johnson , Carl Hu
show 2 more authors
Kaushik Rangadurai Luke Simon
This is my paper
classification cs.IR cs.LG
keywords soft tokensLLM-based recommendationtwo-tower retrievalQ-FormerFourier feature encodingfeature fusionsequential recommendationretrieval
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.

This paper argues that large language models used as recommender backbones do not have to be limited to discrete text tokens. It proposes representing continuous numerical features and dense embedding vectors as soft tokens, which are then inserted directly into the LLM's input embedding sequence. The method is instantiated in a shared-parameter two-tower retrieval model, where the two soft-token streams are refined through an interaction-based fusion module before entering the final LLM tower. On the Beauty, Sports and Outdoors, and Toys and Games benchmarks, the paper reports that interaction-based fusion improves retrieval over LLM-based baselines, and that it consistently outperforms simply concatenating the soft tokens. A sympathetic reading: LLM recommenders can consume the same heterogeneous, continuous signals that classical recommender systems use, provided those signals are first compressed and cross-fused rather than appended without coordination.

Core claim

The central claim is that non-textual recommendation features—numerical scalars such as price, and dense embeddings derived from a richer textual prompt—can be converted into fixed-length soft tokens and consumed by an LLM through its standard token interface, and that doing so improves retrieval quality over a text-only LLM two-tower model. The paper further claims that after this conversion, an explicit interaction stage between the embedding soft tokens and the numerical soft tokens is necessary: direct concatenation of the two heterogeneous streams is not consistently beneficial and can even hurt, whereas a lightweight Transformer fusion module with learnable attention direction produces

What carries the argument

The core mechanism is the soft token: a continuous vector inserted into the LLM's input embedding sequence rather than produced by vocabulary lookup. To build these soft tokens from variable-length feature sequences, the paper uses a Q-Former adapter—a Transformer with a small set of learnable query tokens that cross-attend over the input feature sequence, compressing it to a fixed number of feature-aware tokens. Numerical price values are first mapped to a high-dimensional Fourier feature encoding, then projected and passed through a second Q-Former. The two soft-token streams are then fed into a lightweight Transformer fusion module whose attention mask can be bidirectional or one-way (emb

Load-bearing premise

The paper identifies the gain with soft-token tokenization, but the embedding stream is built from a richer text prompt than the text-only baseline receives, so the added text, not the soft tokens, could be what drives the improvement.

What would settle it

Run the two-tower LLM with the text-only baseline replaced by the same rich prompt used for the embedding stream (either fed directly as text tokens or after simple pooling); if the retrieval gap between text-only and soft-token fusion disappears, soft-token tokenization is not the active ingredient.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • LLM-based recommender systems can incorporate continuous signals (prices, ratings, dense encodings) without verbalizing them into text, preserving scale sensitivity and fine-grained variation.
  • Explicit pre-interaction of heterogeneous soft-token streams is a better default than merging them by concatenation, a point the ablations support across all three datasets.
  • The soft-token interface keeps the LLM's standard token-based input, so existing two-tower retrieval serving (precomputed item embeddings, ANN search) remains intact.
  • The advantage is not free: the embedding stream requires an auxiliary, larger LLM encoder to produce the dense features, plus adapters and a fusion module, so compute and storage trade-offs are part of the deployment picture.
  • Because item-side soft tokens can be precomputed and indexed, the approach remains scalable for candidate generation at serving time.

Where Pith is reading between the lines

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

  • A controlled comparison feeding the text-only baseline the same rich prompt used to build the embedding stream would reveal whether the gain comes from soft-token tokenization or simply from the additional text that the rich prompt contains.
  • The same soft-token interface could be extended to categorical IDs or to learned user/item embeddings from existing classical recommenders, converting legacy features into LLM-visible tokens.
  • The hyperparameter study indicates that the optimal attention direction of the fusion module varies by dataset (bidirectional on Beauty, embedding-to-numerical on the other two), suggesting that the 'interaction' component may require per-dataset tuning rather than a single universal design.
  • Because the numerical stream in this paper is limited to price, a natural test is whether other continuous features—age, delivery time, user engagement counts—obey the same Fourier-plus-Q-Former recipe.

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

4 major / 4 minor

Summary. The paper proposes a soft-token fusion framework that converts continuous numerical features (price) and dense embedding features (encoded from rich prompts by an auxiliary LLM) into soft tokens consumable by a shared-parameter LLM-based two-tower retrieval model. Numerical features are Fourier-encoded and compressed by a numerical Q-Former; embedding features are obtained by pooling the auxiliary encoder's hidden states and compressing them with an embedding Q-Former. A lightweight Transformer fusion module pre-interacts the two soft-token streams before insertion into the final Qwen3-0.6B tower. Experiments on three Amazon benchmarks compare text-only, single-source, concatenation, and interaction-based fusion variants, and compare the interaction variant with published baselines. The paper claims that soft-token fusion improves retrieval performance over LLM-based baselines and that interaction-based fusion is more effective than direct concatenation.

Significance. If the central claims held, the paper would provide a practical interface for integrating non-textual features into LLM-based recommenders without abandoning the token interface, and the interaction-versus-concatenation finding would be a useful architectural insight. The paper has concrete strengths: it uses public Amazon benchmarks, a held-out leave-one-out protocol, a clear two-tower instantiation, and detailed appendices documenting prompt construction and design ablations. The interaction-versus-concatenation comparison is internally consistent across all three datasets. However, the main ablation is confounded by prompt granularity and auxiliary encoder capacity, and the 'strongest overall performance' conclusion is contradicted by the paper's own baseline tables. These issues currently limit the significance of the empirical claims.

major comments (4)
  1. [Appendix A / Eq. (8) / Figure 1] The headline claim 'soft-token fusion improves retrieval performance over LLM-based baselines' is not identifiable from the reported ablations. The embedding stream in Eq. (8) encodes the rich prompt (Appendix A.2) with Qwen3-4B, while the text-only baseline in Tables 2–4 receives only the summarized prompt (Appendix A.1). Thus adding the embedding stream changes three variables at once: soft-token tokenization, the amount of textual metadata, and the capacity of the auxiliary encoder. Figure 1's caption ('the same prompt is used in three ways') is directly contradicted by Appendix A's two distinct prompt granularities. To support the claim, the authors need a control that feeds the rich prompt as ordinary text to the 0.6B tower without soft-token construction, or an embedding stream computed from the summarized prompt only, so that the effect of soft-token tokenization is isolated.
  2. [Conclusion / Tables 5–6] The conclusion that 'interaction-based fusion achieves the strongest overall performance across Beauty, Sports and Outdoors, and Toys and Games' is contradicted by the paper's own baseline tables. On Sports and Outdoors (Table 6), TIGER has higher R@5 (0.0264 vs 0.0253) and higher NDCG@5 and NDCG@10 (0.0181 vs 0.0154 and 0.0225 vs 0.0205). On Beauty (Table 5), TIGER has higher NDCG@5 (0.0321 vs 0.0278) and slightly higher NDCG@10 (0.0384 vs 0.0377). The claim should be qualified to 'competitive' or restricted to R@10 and the Toys dataset, where the proposed model is indeed strongest.
  3. [Section 5.3 / Appendix C] The comparison with baselines is inflated by per-dataset selection of the fusion configuration. Appendix C reports a configuration search on each dataset (e.g., bidirectional 2 layers on Beauty, embedding→numerical 1 layer on Sports, embedding→numerical 2 layers on Toys), and Section 5.3 then reports 'Ours (Interaction)' in Tables 5–7 without stating whether the selection was made on the validation split or on the test set. The leave-one-out protocol defines a validation item, but the paper does not describe using it for hyperparameter selection. The concatenation baseline in Tables 2–4 is not given the same tuning budget. The authors should specify the selection procedure or report a fixed configuration and a tuned concatenation baseline.
  4. [Tables 2–7] No measure of variance or statistical significance is reported. Several key differences are small in absolute terms (e.g., Beauty R@5: 0.0459 vs 0.0454 for TIGER; Sports R@10: 0.0410 vs 0.0400), and no random seeds are given. At minimum, the authors should report mean and standard deviation over multiple seeds, or a paired significance test, for the main comparisons: soft-token vs text-only, and interaction vs concatenation.
minor comments (4)
  1. [Figure 1 / Appendix A] The Figure 1 caption should be revised to reflect that the text, embedding, and numerical streams use different prompt granularities. As written, 'the same prompt is used in three ways' is misleading and contradicts Appendix A.
  2. [Eq. (7) / Eq. (11)] The Fourier encoding definition is duplicated. Define once in Section 3.3 and refer back from Section 4.2.
  3. [References] TIGER is cited twice as Rajput et al. 2023a and 2023b with identical titles and venue. Use a single reference.
  4. [Table 2] The first row uses 'N/A' in the Embedding and Numerical columns while later tables use check marks. Make the formatting consistent across Tables 2–4.

Circularity Check

0 steps flagged

No significant circularity: the central comparisons are held-out empirical ablations; the prompt-granularity and config-selection issues are confounds, not definitional reductions.

full rationale

The paper's derivation chain is empirical rather than definitional. Equations (8)-(18) define a feature-encoding and fusion pipeline: auxiliary LLM encoding, pooling, Q-Former compression, Fourier encoding, transformer fusion, and final LLM encoding. No fitted target or evaluation metric enters these equations, and no equation reduces to a previously fitted value. The central claims — that soft-token fusion helps and that interaction-based fusion beats concatenation — are supported by leave-one-out held-out evaluations against published baselines and ablations. The interaction-vs-concatenation comparison is especially clean because both arms consume the same embedding and numerical soft-token streams, differing only in the pre-interaction module. The prompt-granularity asymmetry (rich prompt for the embedding stream versus summarized prompt for the text-only baseline) is a genuine identifiability threat to the interpretation of the ablation, and the per-dataset selection of the best fusion configuration from Tables 8-10 is a model-selection concern, but neither is circular: the reported gains are not equal to the inputs by construction, and no parameter is fitted to the test labels and then renamed as a prediction. The paper also does not rely on a load-bearing self-citation chain or an imported uniqueness theorem; the Q-Former, Fourier encoding, and TIGER protocol are standard external results used as components, not as evidence that the paper's own conclusions are forced. Therefore, under the stated rules, there is no qualifying circular step.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

The central claim depends on trained model parameters and hand-chosen hyperparameters; there are no new physical entities. The paper's conclusion is empirical rather than a derivation, so the free-parameter count reflects architecture choices rather than fitted physical constants.

free parameters (6)
  • Q-Former query tokens (Qe, Qn) = learned during training; counts Ne/Nn not reported
    Learnable queries define the soft-token bottleneck; the central claim depends on their training.
  • Fourier frequency bands {b_r} = not reported
    Frequency bands for price encoding are chosen by hand and affect numerical soft tokens.
  • Fusion Transformer parameters psi = learned
    The interaction-based fusion module parameters are trained on the data.
  • Fusion attention direction and layer count = dataset-dependent (e.g., bidirectional 2 layers for Beauty; embedding-to-numerical 1 layer for Sports; embedding-to-nume
    Selected per dataset from validation; the main results report the best configuration.
  • Soft-token counts Ne/Nn and pooling stride G = not reported
    Control soft-token compression; not ablated in the main text.
  • Contrastive temperature tau = not reported
    Scaling in Eq. (4); not ablated.
axioms (4)
  • domain assumption Textual prompts and LLM embeddings encode enough semantic signal for sequential recommendation
    The whole approach assumes Qwen3-4B embeddings of rich prompts and the Qwen3-0.6B text tower are useful for next-item retrieval.
  • domain assumption Leave-one-out next-item prediction on Amazon subsets is a valid proxy for recommendation retrieval performance
    The evaluation protocol is borrowed from TIGER; results may not transfer to large-scale deployment.
  • domain assumption Fourier encoding preserves price information that a linear projection would lose
    Motivates Eq. (11); supported only by ablation on three datasets.
  • standard math Standard Transformer attention and Q-Former operations behave as specified
    Equations (5)-(6) and (16) rely on standard attention and feed-forward operations.

pith-pipeline@v1.3.0-alltime-deepseek · 13295 in / 11891 out tokens · 103234 ms · 2026-08-04T01:39:26.736253+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Tokenizing Numerical and Embedding Features for LLM RecSys." pith.science (2026). https://pith.science/paper/6VG6X7P6

@misc{pith2026260710016,
  author       = {Pith},
  title        = {Pith review of: Tokenizing Numerical and Embedding Features for LLM RecSys},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6VG6X7P6}},
  note         = {Machine review of arXiv:2607.10016}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Large language models (LLMs) are increasingly used as backbone architectures for recommender systems because of their strong sequence modeling and representation learning capabilities. However, most LLM-based recommenders operate primarily on discrete textual tokens, whereas practical recommendation pipelines also rely on continuous numerical features and dense embedding features produced by upstream feature engineering or pretrained encoders. This mismatch limits the ability of LLM-based models to exploit fine-grained non-textual signals. We propose a soft-token fusion framework that maps numerical and embedding features into the LLM embedding space, allowing heterogeneous recommendation signals to be consumed through the standard token interface. We instantiate the framework in a shared-parameter LLM-based two-tower retrieval model and introduce an interaction-based fusion module that refines embedding and numerical soft tokens before they are inserted into the final LLM input. Experiments on three Amazon recommendation benchmarks show that soft-token fusion improves retrieval performance over LLM-based baselines, and that interaction-based fusion is more effective than direct concatenation of heterogeneous soft tokens.

Figures

Figures reproduced from arXiv: 2607.10016 by Anil Ramakrishna, Ankit Peshin, Carl Hu, Chiyu Zhang, Feng Qi, Johnson Lui, Justin Johnson, Kaushik Rangadurai, Luke Simon, Zhe Xu.

Figure 1
Figure 1. Figure 1: Overview of the proposed soft-token fusion framework. Only the user tower is [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗

Review history (3 revisions) →

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

48 extracted references · 7 linked inside Pith

  1. [1]

    Proceedings of the 16th ACM Conference on Recommender Systems , pages =

    Recommendation as Language Processing (RLP): A Unified Pretrain, Personalized Prompt & Predict Paradigm (P5) , author =. Proceedings of the 16th ACM Conference on Recommender Systems , pages =. 2022 , doi =

    2022

  2. [2]

    Proceedings of the 17th ACM Conference on Recommender Systems , pages =

    TALLRec: An Effective and Efficient Tuning Framework to Align Large Language Model with Recommendation , author =. Proceedings of the 17th ACM Conference on Recommender Systems , pages =. 2023 , doi =

    2023

  3. [3]

    Proceedings of the 1st Workshop on Deep Learning for Recommender Systems , pages =

    Wide & Deep Learning for Recommender Systems , author =. Proceedings of the 1st Workshop on Deep Learning for Recommender Systems , pages =. 2016 , doi =

    2016

  4. [4]

    Proceedings of the 26th International Joint Conference on Artificial Intelligence , pages =

    DeepFM: A Factorization-Machine Based Neural Network for CTR Prediction , author =. Proceedings of the 26th International Joint Conference on Artificial Intelligence , pages =

  5. [5]

    Proceedings of the ADKDD'17 , pages =

    Deep & Cross Network for Ad Click Predictions , author =. Proceedings of the ADKDD'17 , pages =. 2017 , doi =

    2017

  6. [6]

    arXiv preprint arXiv:1906.00091 , year =

    Deep Learning Recommendation Model for Personalization and Recommendation Systems , author =. arXiv preprint arXiv:1906.00091 , year =

    Pith/arXiv arXiv 1906

  7. [7]

    Proceedings of the 10th ACM Conference on Recommender Systems , pages =

    Deep Neural Networks for YouTube Recommendations , author =. Proceedings of the 10th ACM Conference on Recommender Systems , pages =. 2016 , doi =

    2016

  8. [8]

    Proceedings of the 13th ACM Conference on Recommender Systems , pages =

    Sampling-Bias-Corrected Neural Modeling for Large Corpus Item Recommendations , author =. Proceedings of the 13th ACM Conference on Recommender Systems , pages =. 2019 , doi =

    2019

  9. [9]

    Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing , pages =

    Prefix-Tuning: Optimizing Continuous Prompts for Generation , author =. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing , pages =. 2021 , doi =

    2021

  10. [10]

    Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , pages =

    The Power of Scale for Parameter-Efficient Prompt Tuning , author =. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , pages =. 2021 , doi =

    2021

  11. [11]

    Proceedings of the 40th International Conference on Machine Learning , pages =

    BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models , author =. Proceedings of the 40th International Conference on Machine Learning , pages =

  12. [12]

    arXiv preprint arXiv:2305.19860 , year =

    A Survey on Large Language Models for Recommendation , author =. arXiv preprint arXiv:2305.19860 , year =

    Pith/arXiv arXiv

  13. [13]

    arXiv preprint arXiv:2401.04997 , year =

    Tapping the Potential of Large Language Models as Recommender Systems: A Comprehensive Framework and Empirical Analysis , author =. arXiv preprint arXiv:2401.04997 , year =

    Pith/arXiv arXiv

  14. [14]

    Proceedings of the ACM Web Conference 2024 , pages =

    Representation Learning with Large Language Models for Recommendation , author =. Proceedings of the ACM Web Conference 2024 , pages =. 2024 , doi =

    2024

  15. [15]

    Proceedings of the 17th ACM International Conference on Web Search and Data Mining , pages =

    LLMRec: Large Language Models with Graph Augmentation for Recommendation , author =. Proceedings of the 17th ACM International Conference on Web Search and Data Mining , pages =. 2024 , doi =

    2024

  16. [16]

    arXiv preprint arXiv:2307.15780 , year =

    LLM-Rec: Personalized Recommendation via Prompting Large Language Models , author =. arXiv preprint arXiv:2307.15780 , year =

    Pith/arXiv arXiv

  17. [17]

    arXiv preprint arXiv:2309.01157 , year =

    Large Language Models for Generative Recommendation: A Survey and Visionary Discussions , author =. arXiv preprint arXiv:2309.01157 , year =

    Pith/arXiv arXiv

  18. [18]

    Advances in Neural Information Processing Systems , volume =

    Recommender Systems with Generative Retrieval , author =. Advances in Neural Information Processing Systems , volume =

  19. [19]

    Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages =

    LLaRA: Large Language-Recommendation Assistant , author =. Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages =. 2024 , doi =

    2024

  20. [20]

    Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , pages =

    xDeepFM: Combining Explicit and Implicit Feature Interactions for Recommender Systems , author =. Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , pages =

  21. [21]

    Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , pages =

    Deep Interest Network for Click-Through Rate Prediction , author =. Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , pages =

  22. [22]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume =

    Deep Interest Evolution Network for Click-Through Rate Prediction , author =. Proceedings of the AAAI Conference on Artificial Intelligence , volume =. 2019 , doi =

    2019

  23. [23]

    Proceedings of the 13th ACM Conference on Recommender Systems , pages =

    FiBiNET: Combining Feature Importance and Bilinear Feature Interaction for Click-Through Rate Prediction , author =. Proceedings of the 13th ACM Conference on Recommender Systems , pages =

  24. [24]

    Proceedings of the 28th ACM International Conference on Information and Knowledge Management , pages =

    AutoInt: Automatic Feature Interaction Learning via Self-Attentive Neural Networks , author =. Proceedings of the 28th ACM International Conference on Information and Knowledge Management , pages =

  25. [25]

    arXiv preprint arXiv:2012.06678 , year =

    TabTransformer: Tabular Data Modeling Using Contextual Embeddings , author =. arXiv preprint arXiv:2012.06678 , year =

    Pith/arXiv arXiv 2012

  26. [26]

    Advances in Neural Information Processing Systems , year =

    Revisiting Deep Learning Models for Tabular Data , author =. Advances in Neural Information Processing Systems , year =

  27. [27]

    Advances in Neural Information Processing Systems , year =

    Flamingo: A Visual Language Model for Few-Shot Learning , author =. Advances in Neural Information Processing Systems , year =

  28. [28]

    Advances in Neural Information Processing Systems , volume =

    Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional Domains , author =. Advances in Neural Information Processing Systems , volume =

  29. [29]

    European Conference on Computer Vision , pages =

    NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis , author =. European Conference on Computer Vision , pages =

  30. [30]

    Advances in Neural Information Processing Systems , year =

    Attention Is All You Need , author =. Advances in Neural Information Processing Systems , year =

  31. [31]

    Advances in Neural Information Processing Systems , year =

    ViLBERT: Pretraining Task-Agnostic Visiolinguistic Representations for Vision-and-Language Tasks , author =. Advances in Neural Information Processing Systems , year =

  32. [32]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , year =

    Multimodal Learning with Transformers: A Survey , author =. IEEE Transactions on Pattern Analysis and Machine Intelligence , year =

  33. [33]

    Proceedings of the 25th International Conference on World Wide Web , series =

    Ups and Downs: Modeling the Visual Evolution of Fashion Trends with One-Class Collaborative Filtering , author =. Proceedings of the 25th International Conference on World Wide Web , series =. 2016 , publisher =

    2016

  34. [34]

    arXiv preprint arXiv:2505.09388 , year =

    Qwen3 Technical Report , author =. arXiv preprint arXiv:2505.09388 , year =

    Pith/arXiv arXiv

  35. [35]

    International Conference on Learning Representations , year =

    Session-based Recommendations with Recurrent Neural Networks , author =. International Conference on Learning Representations , year =

  36. [36]

    Proceedings of the Eleventh ACM International Conference on Web Search and Data Mining , pages =

    Personalized Top-N Sequential Recommendation via Convolutional Sequence Embedding , author =. Proceedings of the Eleventh ACM International Conference on Web Search and Data Mining , pages =. 2018 , publisher =

    2018

  37. [37]

    Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , pages =

    Hierarchical Gating Networks for Sequential Recommendation , author =. Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , pages =. 2019 , publisher =

    2019

  38. [38]

    2018 IEEE International Conference on Data Mining , pages =

    Self-Attentive Sequential Recommendation , author =. 2018 IEEE International Conference on Data Mining , pages =. 2018 , publisher =

    2018

  39. [39]

    2019 , publisher =

    Sun, Fei and Liu, Jun and Wu, Jian and Pei, Changhua and Lin, Xiao and Ou, Wenwu and Jiang, Peng , booktitle =. 2019 , publisher =

    2019

  40. [40]

    Proceedings of the 28th International Joint Conference on Artificial Intelligence , pages =

    Feature-level Deeper Self-Attention Network for Sequential Recommendation , author =. Proceedings of the 28th International Joint Conference on Artificial Intelligence , pages =. 2019 , doi =

    2019

  41. [41]

    2020 , publisher =

    Zhou, Kun and Wang, Hui and Zhao, Wayne Xin and Zhu, Yutao and Wang, Sirui and Zhang, Fuzheng and Wang, Zhongyuan and Wen, Ji-Rong , booktitle =. 2020 , publisher =

    2020

  42. [42]

    2022 IEEE 38th International Conference on Data Engineering , pages =

    Contrastive Learning for Sequential Recommendation , author =. 2022 IEEE 38th International Conference on Data Engineering , pages =. 2022 , publisher =

    2022

  43. [43]

    Proceedings of the Fifteenth ACM International Conference on Web Search and Data Mining , pages =

    Contrastive Learning for Representation Degeneration Problem in Sequential Recommendation , author =. Proceedings of the Fifteenth ACM International Conference on Web Search and Data Mining , pages =. 2022 , publisher =

    2022

  44. [44]

    Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages =

    Lighter and Better: Low-Rank Decomposed Self-Attention Networks for Next-Item Recommendation , author =. Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages =. 2021 , publisher =

    2021

  45. [45]

    Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages =

    Frequency Enhanced Hybrid Attention Network for Sequential Recommendation , author =. Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages =. 2023 , publisher =

    2023

  46. [46]

    Proceedings of the 13th International Conference on Web Search and Data Mining , pages =

    Time Interval Aware Self-Attention for Sequential Recommendation , author =. Proceedings of the 13th International Conference on Web Search and Data Mining , pages =. 2020 , publisher =

    2020

  47. [47]

    Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages =

    Decoupled Side Information Fusion for Sequential Recommendation , author =. Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages =. 2022 , publisher =

    2022

  48. [48]

    Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages =

    Learning Attribute as Explicit Relation for Sequential Recommendation , author =. Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages =. 2025 , publisher =

    2025