Pith. sign in

REVIEW 3 major objections 5 minor 71 references

Does Knowledge Distillation Matter for Large Language Model based Bundle Generation?

T0 review · 3 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read This paper claims that explicit knowledge distillation lets a small 8-billion-parameter student model generate product bundles as well as—and on precision and coverage better than—a large teacher model, at a fraction of the memory cost…

desk verdict Useful empirical map of explicit KD for LLM-based bundle generation, but the 'student beats teacher' claim rests on a subset-match evaluation and post-hoc selection, so treat it as suggestive, not established. read the letter →

arxiv 2504.17220 v1 pith:RJAH4S7N submitted 2025-04-24 cs.CL cs.IR

Kaidong Feng , Zhu Sun , Jie Yang , Hui Fang , Xinghua Qu , Wenyuan Liu This is my paper · ORCID
classification cs.CLcs.IR
keywords knowledgedistillationbundlegenerationlargelanguagemodelsin-contextlearningsupervisedfine-tuningrecommendersystemsefficiencychain-of-thought
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper asks whether knowledge distillation can make LLM-based bundle generation cheap without losing quality. It distills three increasingly rich kinds of knowledge—frequent category patterns, formalized rules, and chain-of-thought "deep thoughts"—from GPT-3.5-turbo into Llama3.1-8B, then studies how the format, quantity, and method of using that knowledge change performance. Across three product-bundle datasets, the best student models match or outperform the teacher on Precision and Coverage in every domain while using far less memory, though they consistently fall short on Recall. The paper concludes that distillation works, that supervised fine-tuning with distilled knowledge is the most reliable way to inject it, and that how the knowledge is used matters more than what form it takes.

What carries the argument

The load-bearing mechanism is a three-level progressive knowledge distillation pipeline. Frequent patterns (category co-occurrences mined with the Apriori algorithm) capture what items belong together; formalized rules (generated by having the teacher reflect on its own bundling errors) capture when and why items should be grouped; deep thoughts (produced by chain-of-thought prompting) capture context-sensitive reasoning about user intent. These knowledge artifacts are delivered to the student either as retrieved prompt context (in-context learning) or as training signals (supervised fine-tuning with QLoRA), alone or combined. The framework treats bundle generation as producing JSON from a session's product list, and scores each generated bundle as a hit if it exactly matches or is a subset of a ground-truth bundle.

What would settle it

Recompute Precision, Recall, and Coverage using only exact bundle matches, or a credit rule that penalizes partial subsets, and check whether the student still beats the teacher on Precision and Coverage in all three domains; the claim fails if the advantage shrinks or reverses.

Watch

Extended reading notes

Core claim

The central claim is that explicit knowledge distillation is a viable route to efficient LLM-based bundle generation: a student model trained on teacher-extracted knowledge can reach performance comparable to, and on Precision and Coverage even superior to, the large teacher model, while cutting memory demand by roughly an order of magnitude. This is established empirically across Electronic, Clothing, and Food datasets, with the best students beating GPT-3.5-turbo on Precision and Coverage in every domain and exceeding the AICL baseline on Coverage (and on Precision in Food). The paper also claims a nuanced dependency story: supervised fine-tuning benefits more consistently than in-context learning, in-context learning improves with more distilled knowledge while fine-tuning peaks around a 70 percent sampling ratio, cross-domain and multi-format aggregation helps fine-tuning, and the utilization method has the largest overall effect on performance.

Load-bearing premise

The evaluation counts a generated bundle as correct if it completely matches or is a subset of a ground-truth bundle, so a student that emits many small valid subsets can inflate Precision and Coverage without ever producing the full bundles—and the paper's "student surpasses teacher" result inherits this credit assignment.

Editorial extensions

If this is right

  • An 8-billion-parameter student fine-tuned on distilled knowledge can serve bundle generation at roughly one-tenth the memory footprint of a large API teacher, with better precision and coverage but a consistent recall penalty.
  • Practitioners should expect in-context learning to improve as the knowledge pool grows, while fine-tuning peaks at an intermediate amount of knowledge; more data is not always better for training the student.
  • The choice of how knowledge is used is the strongest lever: combining fine-tuning with in-context learning is generally best when the knowledge types at the two stages are coherent, and fine-tuning alone is more reliable than in-context learning alone.
  • Accumulating knowledge across domains and formats helps a fine-tuned student generalize, but gives little benefit to an in-context-learning student, which only retrieves knowledge similar to the current session.
  • Among the three factors studied, knowledge format matters least; effort spent on how distilled knowledge is applied pays more than effort spent on which textual format it takes.

Reading between the lines

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

  • Because hits include subsets of ground-truth bundles, the reported Precision and Coverage gains may partly reflect students learning to emit many small valid combinations rather than reproducing full bundles; an exact-match evaluation could materially change the teacher-student ranking.
  • The efficiency comparison uses memory footprint rather than wall-clock latency, so a fair latency benchmark would be the decisive test for real-time recommendation deployment.
  • The recall deficit suggests explicit textual distillation captures common regularities but loses the teacher's coverage of rare or context-specific bundles, making implicit distillation of internal representations or adaptive knowledge selection a natural next step.
  • A testable extension would apply the same format-quantity-utilization grid to other structured generation tasks, such as itinerary or shopping-list generation, where subset scoring is not used.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. This paper studies whether explicit knowledge distillation from a large teacher LLM (GPT-3.5-turbo) improves bundle generation by a smaller student model (Llama3.1-8B). It proposes a framework with three knowledge formats (frequent patterns, formalized rules, deep thoughts), four ways of controlling knowledge quantity (random/length/diversity/difficulty sampling at ratios 0.1-1.0, plus domain and format accumulation), and three utilization methods (ICL, SFT, ICL+SFT). Experiments on Electronic, Clothing, and Food datasets compare student variants with conventional baselines and with the teacher, and report that the best student models outperform the teacher in Precision and Coverage while lagging in Recall, at much lower training/inference resource estimates. The paper also analyzes which factor (format, quantity, utilization) has the largest impact on performance.

Significance. The study is a useful systematic ablation in an underexplored combination of LLMs, knowledge distillation, and bundle generation. Its strengths are the transparent three-factor design, the use of three publicly available datasets, the detailed hyperparameter search, and the explicit reporting of resource estimates. If the headline result survived a stricter evaluation, the practical implication (a cheaper bundle generator that beats the teacher on Precision and Coverage) would be valuable. At present, however, the headline rests on an evaluation rule that rewards partial subsets and on post hoc selection of the best configuration, so the empirical claims should be treated as conditional until re-analyzed.

major comments (3)
  1. [Section 4.1; Tables 5-6; Section 4.5.1] The definition of a hit bundle as one that 'completely matches or is a subset of a ground truth bundle' is load-bearing for the central claim. Because Precision, Recall, and Coverage all inherit this rule, a student that emits many small valid subsets receives full credit for Precision and Coverage without reconstructing any ground-truth bundle, while the teacher, which tends to produce larger bundles, is penalized more harshly for a single wrong item. The pattern in Table 6 (Precision gains of 12-17% coupled with Recall losses of 16-17%) is exactly what this mechanism predicts. Please re-run the evaluation with an exact-match definition of a hit, report both subset-match and exact-match results, and include the average size of predicted bundles per model so readers can assess whether the student's advantage is an artifact of emitting smaller bundles.
  2. [Section 4.5.1; Table 6] The main comparison selects, after seeing the results, the best-performing student configuration per domain and metric among the RQ1/RQ2/RQ3 variants, and then reports its improvement over the teacher. No repeated runs or confidence intervals are provided, and the only significance test in the paper (the paired t-test in Section 4.3.1) addresses a different question, namely whether SFT is more sensitive than ICL to sampling strategy. The claim that students 'consistently outperform the teacher' should be supported with mean and standard deviation over multiple seeds, significance tests with a multiple-comparison correction for the Table 6 selections, or explicitly labeled as an exploratory best-case analysis.
  3. [Section 4.3.1] The paired t-test reported in Section 4.3.1 pools standard deviations computed across 'different sample ratios, knowledge formats, and domains' as if they were independent replicates, but these quantities come from overlapping runs and are not independent. The test should be re-run on independent repeated runs or omitted; as written it does not provide statistical support for the claim that sampling strategy has a greater impact on SFT than on ICL.
minor comments (5)
  1. [Section 3.1.3] In the prompt text, 'A void common mistakes' should read 'Avoid common mistakes'.
  2. [Section 4.5.2] The memory figures for the teacher (e.g., 'around 1450GB for full fine-tuning') are stated without a measurement protocol, and GPT-3.5-turbo is API-only; please clarify whether these are estimates and how they were obtained.
  3. [Section 4.1] Precision, Recall, and Coverage are described verbally but no formulas are given; adding explicit equations would make the subset-match rule and its effect on the results easier to audit.
  4. [Section 4.1] The paper does not report how many model outputs failed JSON parsing or how such failures were handled; since prompts require JSON-only output, a parse-failure rate should be given for each model and dataset.
  5. [Reproducibility] No code or data release is mentioned; providing implementation details such as prompts, retrieval scripts, and hyperparameter configurations would improve reproducibility.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the knowledge-distillation comparisons are empirical and self-contained, with self-citations to the authors' datasets, metrics, and AICL baseline serving as contextual benchmarks rather than load-bearing reductions.

full rationale

This paper is an empirical study rather than a formal derivation, so the primary circularity failure modes do not apply. The distilled knowledge (patterns, rules, thoughts) is extracted from training sessions, while student models are evaluated on held-out test sessions, so there is no direct leakage of the evaluation target into the training signal. The 'student surpasses teacher on Precision and Coverage' claim is obtained by comparing actual model outputs under a stated metric, not by renaming a fitted parameter as a prediction. The authors' self-citations to their own datasets [47, 50], to the evaluation metrics adopted from [47, 50], and to the AICL baseline [48] are normal uses of prior public resources and are not used to define away the comparison. The subset-match credit rule in Section 4.1, which counts a generated bundle as a hit if it is a subset of a ground-truth bundle, is a legitimate validity concern: it may inflate student Precision and Coverage if students emit many small partial bundles, and the accompanying Recall drops in Table 6 are consistent with that mechanism. However, this is a methodological and interpretational threat, not a circular derivation: the metric is not derived from the conclusion, and the student outputs are not constructed by the paper to satisfy the metric. Similarly, selecting the best-performing student configuration across RQ1-RQ3 variants is a selection-bias concern rather than circular reasoning. Overall, the central empirical comparison is self-contained against external models and held-out data, so the circularity score is low.

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

The central empirical claims rest on dataset labels, the subset-based hit metric, the sufficiency of explicit textual knowledge, and several hand-set thresholds; none of these are independently established in the paper beyond the reported experiments.

free parameters (4)
  • BERT similarity threshold for rule/thought deduplication = 0.8
    Section 3.2.1 uses a cosine similarity threshold of 0.8 to filter redundant distilled rules and thoughts; this hand-set value directly changes the measured quantity of distilled knowledge for RQ2.
  • Apriori minimum support threshold = not reported
    Section 3.1.2 applies Apriori to mine frequent patterns but does not report min support or confidence; without these, pattern sets and their quantities are not reproducible.
  • QLoRA hyperparameters = learning rate in {2e-5, 8e-5, 2e-4}, epochs {3,4,5}, rank {8,16,32}, alpha {8,16,32}
    Section 4.1 selects these via grid search; results depend on these choices, though they are standard tuning parameters.
  • Length-based sampling group boundaries = session lengths [2-4], [5-7], [8-10]
    Section 3.2.1 uses hand-chosen bins for length-based sampling; arbitrary binning can influence which knowledge is distilled and the measured quantity-performance relationship.
assumptions (6)
  • domain assumption Sessions contain ground-truth bundles with user intents, and bundle generation is defined as identifying those bundles within each session.
    Section 3.1.1 relies on dataset labels from the authors' previous SIGIR 2022 paper; the entire evaluation treats these labels as ground truth.
  • domain assumption A generated bundle counts as correct if it fully matches or is a subset of a ground-truth bundle.
    Section 4.1 defines hit bundles this way; Precision, Recall, and Coverage all inherit this subset-based credit, which can favor models that emit many small bundles.
  • ad hoc to paper Explicit textual knowledge (patterns, rules, thoughts) is sufficient to transfer a meaningful portion of teacher competence to the student.
    The whole KD framework assumes this; the Limitations section concedes implicit knowledge in hidden states is not captured, so the assumption is acknowledged by the authors rather than demonstrated.
  • ad hoc to paper Filtering rules and thoughts by BERT semantic similarity with threshold 0.8 preserves useful knowledge while removing redundancy.
    Section 3.2.1 assumes this threshold; no validation is provided for whether discarded items are truly redundant.
  • domain assumption Varying the sampled data portion yields different amounts of distilled knowledge.
    Section 3.2.1 states and then verifies this with Figure 3, so it is an explicit assumption rather than a derived result.
  • domain assumption GPT-3.5-turbo is a suitable teacher for bundle-generation knowledge.
    Section 4.1 selects GPT-3.5-turbo as teacher without comparing other teachers; all distilled knowledge inherits this choice.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Does Knowledge Distillation Matter for Large Language Model based Bundle Generation?." pith.science (2026). https://pith.science/paper/RJAH4S7N

@misc{pith2026250417220,
  author       = {Pith},
  title        = {Pith review of: Does Knowledge Distillation Matter for Large Language Model based Bundle Generation?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RJAH4S7N}},
  note         = {Machine review of arXiv:2504.17220}
}
read the original abstract

LLMs are increasingly explored for bundle generation, thanks to their reasoning capabilities and knowledge. However, deploying large-scale LLMs introduces significant efficiency challenges, primarily high computational costs during fine-tuning and inference due to their massive parameterization. Knowledge distillation (KD) offers a promising solution, transferring expertise from large teacher models to compact student models. This study systematically investigates knowledge distillation approaches for bundle generation, aiming to minimize computational demands while preserving performance. We explore three critical research questions: (1) how does the format of KD impact bundle generation performance? (2) to what extent does the quantity of distilled knowledge influence performance? and (3) how do different ways of utilizing the distilled knowledge affect performance? We propose a comprehensive KD framework that (i) progressively extracts knowledge (patterns, rules, deep thoughts); (ii) captures varying quantities of distilled knowledge through different strategies; and (iii) exploits complementary LLM adaptation techniques (in-context learning, supervised fine-tuning, combination) to leverage distilled knowledge in small student models for domain-specific adaptation and enhanced efficiency. Extensive experiments provide valuable insights into how knowledge format, quantity, and utilization methodologies collectively shape LLM-based bundle generation performance, exhibiting KD's significant potential for more efficient yet effective LLM-based bundle generation.

Figures

Figures reproduced from arXiv: 2504.17220 by the authors.

Figure 1
Figure 1. Example bundles for (1) a camera and its accessories; and (2) mystery, thriller, and historical fiction [ [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The overview of our proposed knowledge distillation framework. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Variation in knowledge quantity as sampling ratio increases (Pattern: 1st Column; Rule: 2nd Column; Thought: 3rd Column). [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: Performance comparison of Llama3.1-ICL under different sampling strategies and ratios in the Electronic domain. [PITH_FULL_IMAGE:figures/full_fig_p015_4.png]
Figure 5
Figure 5. Figure 5: Performance comparison of Llama3.1-ICL under different sampling strategies and ratios in the Clothing domain. [PITH_FULL_IMAGE:figures/full_fig_p016_5.png]
Figure 6
Figure 6. Figure 6: Performance comparison of Llama3.1-ICL under different sampling strategies and ratios in the Food domain. [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]
Figure 7
Figure 7. Figure 7: Performance comparison of Llama3.1-SFT under different sampling strategies and ratios in the Electronic domain. [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]
Figure 8
Figure 8. Figure 8: Performance comparison of Llama3.1-SFT under different sampling strategies and ratios in the Clothing domain. [PITH_FULL_IMAGE:figures/full_fig_p017_8.png]
Figure 9
Figure 9. Figure 9: Performance comparison of Llama3.1-SFT under different sampling strategies and ratios in the Food domain. [PITH_FULL_IMAGE:figures/full_fig_p018_9.png]
Figure 10
Figure 10. Figure 10: The results of knowledge accumulation from different domains in the Electronic, Clothing, and Food domains. [PITH_FULL_IMAGE:figures/full_fig_p019_10.png]
Figure 11
Figure 11. Figure 11: Results of knowledge accumulation from different formats in the Electronic, Clothing, and Food domains using Llama3.1-ICL. [PITH_FULL_IMAGE:figures/full_fig_p020_11.png]
Figure 12
Figure 12. Figure 12: Results of knowledge accumulation from different formats in the Electronic, Clothing, and Food domains using Llama3.1-SFT. [PITH_FULL_IMAGE:figures/full_fig_p020_12.png]
Figure 13
Figure 13. Figure 13: Relative performance differences of various knowledge utilization strategies on student LLMs across the three domains. [PITH_FULL_IMAGE:figures/full_fig_p021_13.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

71 extracted references · 66 canonical work pages

  1. [1]

    Rakesh Agrawal, Ramakrishnan Srikant, et al. 1994. Fast algorithms for mining association rules. In Proceedings of 20th International Conference on Very Large Scale Data Bases (VLDB) , Vol. 1215. 487–499

    work page 1994

  2. [2]

    Tzoof Avny Brosh, Amit Livne, Oren Sar Shalom, Bracha Shapira, and Mark Last. 2022. BRUCE: Bundle recommendation using contextualized item embeddings. In Proceedings of the 16th ACM Conference on Recommender Systems (RecSys) . 237–245

    work page 2022

  3. [3]

    Jinze Bai, Chang Zhou, Junshuai Song, Xiaoru Qu, Weiting An, Zhao Li, and Jun Gao. 2019. Personalized bundle list recommendation. In The Web Conference (TheWebConf). 60–71

    work page 2019

  4. [4]

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

    work page 2023

  5. [5]

    Moran Beladev, Lior Rokach, and Bracha Shapira. 2016. Recommender systems for product bundling. Knowledge-Based Systems (KBS) 111 (2016), 193–206

    work page 2016

  6. [6]

    Da Cao, Liqiang Nie, Xiangnan He, Xiaochi Wei, Shunzhi Zhu, and Tat-Seng Chua. 2017. Embedding factorization models for jointly recommending items and user generated lists. In Proceedings of the 40th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR). 585–594

    work page 2017

  7. [7]

    Jianxin Chang, Chen Gao, Xiangnan He, Depeng Jin, and Yong Li. 2020. Bundle recommendation with graph convolutional networks. In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR) . 1673–1676

    work page 2020

  8. [8]

    Jianxin Chang, Chen Gao, Xiangnan He, Depeng Jin, and Yong Li. 2021. Bundle recommendation and generation with graph neural networks. IEEE Transactions on Knowledge and Data Engineering (TKDE) 35, 3 (2021), 2326–2340

    work page 2021

Show all 71 references
  1. [9]

    Liang Chen, Yang Liu, Xiangnan He, Lianli Gao, and Zibin Zheng. 2019. Matching user with item set: Collaborative bundle recommendation with deep attention network.. In International Joint Conference on Artificial Intelligence (IJCAI) . 2095–2101

    2019

  2. [10]

    Wen Chen, Pipei Huang, Jiaming Xu, Xin Guo, Cheng Guo, Fei Sun, Chao Li, Andreas Pfadler, Huan Zhao, and Binqiang Zhao. 2019. POG: Personalized outfit generation for fashion recommendation at Alibaba iFashion. In Proceedings of the 25th ACM SIGKDD International Conference on K...

    2019

  3. [11]

    Yu Cui, Feng Liu, Pengbo Wang, Bohao Wang, Heng Tang, Yi Wan, Jun Wang, and Jiawei Chen. 2024. Distillation matters: Empowering sequential recommenders to match the performance of large language models. In Proceedings of the 18th ACM Conference on Recommender Systems (RecSys) ...

    2024

  4. [12]

    Sunhao Dai, Ninglu Shao, Haiyuan Zhao, Weijie Yu, Zihua Si, Chen Xu, Zhongxiang Sun, Xiao Zhang, and Jun Xu. 2023. Uncovering chatgpt’s capabilities in recommender systems. In Proceedings of the 17th ACM Conference on Recommender Systems (RecSys) . 1126–1132

    2023

  5. [13]

    Qilin Deng, Kai Wang, Minghao Zhao, Runze Wu, Yu Ding, Zhene Zou, Yue Shang, Jianrong Tao, and Changjie Fan. 2021. Build your own bundle - a neural combinatorial optimization method. In Proceedings of the 29th ACM International Conference on Multimedia (MM) . 2625–2633

    2021

  6. [14]

    Qilin Deng, Kai Wang, Minghao Zhao, Zhene Zou, Runze Wu, Jianrong Tao, Changjie Fan, and Liang Chen. 2020. Personalized bundle recommendation in online games. In Proceedings of the 29th ACM International Conference on Information and Knowledge Management (CIKM) . 2381–2388

    2020

  7. [15]

    Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. 2023. QLORA: Efficient finetuning of quantized llms. In Proceedings of the 37th International Conference on Neural Information Processing Systems (NeurIPS) (NeurIPS ’23)

    2023

  8. [16]

    Paolo Dragone, Giovanni Pellegrini, Michele Vescovi, Katya Tentori, and Andrea Passerini. 2018. No more ready-made deals: Constructive recommendation for telco service bundling. In Proceedings of the 12th ACM Conference on Recommender Systems (RecSys) . 163–171

    2018

  9. [17]

    Yan Fang, Xinyue Xiao, Xiaoyu Wang, and Huiqing Lan. 2018. Customized bundle recommendation by association rules of product categories for online supermarkets. In The 3rd International Conference on Data Science in Cyberspace (DSC) . 472–475. Manuscript submitted to ACM 26 Kai...

    2018

  10. [18]

    Robert Garfinkel, Ram Gopal, Arvind Tripathi, and Fang Yin. 2006. Design of a shopbot and recommender system for bundle purchases. Decision Support Systems (DSS) 42, 3 (2006), 1974–1986

    2006

  11. [19]

    Yong Ge, Hui Xiong, Alexander Tuzhilin, and Qi Liu. 2014. Cost-aware collaborative filtering for travel tour recommendations. ACM Transactions on Information Systems (TOIS) 32, 1 (2014), 1–31

    2014

  12. [20]

    Judy Harris and Edward A Blair. 2006. Consumer preference for product bundles: The role of reduced search costs. Journal of the Academy of Marketing Science 34, 4 (2006), 506–513

    2006

  13. [21]

    Jesse Harte, Wouter Zorgdrager, Panos Louridas, Asterios Katsifodimos, Dietmar Jannach, and Marios Fragkoulis. 2023. Leveraging large language models for sequential recommendation. In Proceedings of the 17th ACM Conference on Recommender Systems (RecSys) . 1096–1102

    2023

  14. [22]

    Ruining He and Julian McAuley. 2016. Ups and downs: Modeling the visual evolution of fashion trends with one-class collaborative filtering. In Proceedings of the 25th International Conference on World Wide Web (WWW) . 507–517

    2016

  15. [23]

    Yun He, Jianling Wang, Wei Niu, and James Caverlee. 2019. A hierarchical self-attentive model for recommending user-generated item lists. In Proceedings of the 28th ACM International Conference on Information and Knowledge Management (CIKM) . 1481–1490

    2019

  16. [24]

    Yun He, Yin Zhang, Weiwen Liu, and James Caverlee. 2020. Consistency-aware recommendation for user-generated item list continuation. In Proceedings of the 13th International Conference on Web Search and Data Mining (WSDM) . 250–258

    2020

  17. [25]

    Zhankui He, Zhouhang Xie, Rahul Jha, Harald Steck, Dawen Liang, Yesu Feng, Bodhisattwa Majumder, Nathan Kallus, and Julian Mcauley. 2023. Large language models as zero-shot conversational recommenders. In Proceedings of the 32nd ACM International Conference on Information and ...

    2023

  18. [26]

    Zhankui He, Handong Zhao, Tong Yu, Sungchul Kim, Fan Du, and Julian McAuley. 2022. Bundle mcr: Towards conversational bundle recommendation. In Proceedings of the 16th ACM Conference on Recommender Systems (RecSys) . 288–298

    2022

  19. [27]

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. 2015. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015)

    2015 arXiv

  20. [28]

    Hyunsik Jeon, Jong-eun Lee, Jeongin Yun, and U Kang. 2024. Cold-start bundle recommendation via popularity-based coalescence and curriculum heating. In Proceedings of the ACM Web Conference 2024 . 3277–3286

    2024

  21. [29]

    Pigi Kouki, Ilias Fountalis, Nikolaos Vasiloglou, Nian Yan, Unaiza Ahsan, Khalifeh Al Jadda, and Huiming Qu. 2019. Product collection recommendation in online retail. In Proceedings of the 13th ACM Conference on Recommender Systems (RecSys) . 486–490

    2019

  22. [30]

    Shiyao Li, Xuefei Ning, Luning Wang, Tengxuan Liu, Xiangsheng Shi, Shengen Yan, Guohao Dai, Huazhong Yang, and Yu Wang. 2024. Evaluating quantized large language models. In Proceedings of the 41st International Conference on Machine Learning (ICML) . Article 1144, 45 pages

    2024

  23. [31]

    Jianghao Lin, Rong Shan, Chenxu Zhu, Kounianhua Du, Bo Chen, Shigang Quan, Ruiming Tang, Yong Yu, and Weinan Zhang. 2024. Rella: Retrieval-enhanced large language models for lifelong sequential behavior comprehension in recommendation. In Proceedings of the ACM Web Conference ...

    2024

  24. [32]

    Guannan Liu, Yanjie Fu, Guoqing Chen, Hui Xiong, and Can Chen. 2017. Modeling buying motives for personalized product bundle recommendation. ACM Transactions on Knowledge Discovery from Data (TKDD) 11, 3 (2017), 1–26

    2017

  25. [33]

    Qi Liu, Yong Ge, Zhongmou Li, Enhong Chen, and Hui Xiong. 2011. Personalized travel package recommendation. In IEEE 11th International Conference on Data Mining (ICDM) . 407–416

    2011

  26. [34]

    Xiaohao Liu, Jie Wu, Zhulin Tao, Yunshan Ma, Yinwei Wei, and Tat-seng Chua. 2025. Fine-tuning multimodal large language models for product bundling. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD) . 848–858

    2025

  27. [35]

    Yidan Liu, Min Xie, and Laks VS Lakshmanan. 2014. Recommending user generated item lists. In Proceedings of the 8th ACM Conference on Recommender Systems (RecSys). 185–192

    2014

  28. [36]

    Yunshan Ma, Yingzhi He, Xiang Wang, Yinwei Wei, Xiaoyu Du, Yuyangzi Fu, and Tat-Seng Chua. 2024. Multicbr: Multi-view contrastive learning for bundle recommendation. ACM Transactions on Information Systems (TOIS) 42, 4 (2024), 1–23

    2024

  29. [37]

    Yunshan Ma, Yingzhi He, An Zhang, Xiang Wang, and Tat-Seng Chua. 2022. CrossCBR: Cross-view contrastive learning for bundle recommendation. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD) . 1233–1241

    2022

  30. [38]

    Yunshan Ma, Yingzhi He, Wenjun Zhong, Xiang Wang, Roger Zimmermann, and Tat-Seng Chua. 2024. CIRP: Cross-item relational pre-training for multimodal product bundling. In Proceedings of the 32nd ACM International Conference on Multimedia (MM) . 9641–9649

    2024

  31. [39]

    Yunshan Ma, Xiaohao Liu, Yinwei Wei, Zhulin Tao, Xiang Wang, and Tat-Seng Chua. 2024. Leveraging multimodal features and item-level user feedback for bundle construction. In Proceedings of the 17th ACM International Conference on Web Search and Data Mining (WSDM) . 510–519

    2024

  32. [40]

    Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, et al. 2023. Self-refine: Iterative refinement with self-feedback. In Advances in Neural Information Processing Systems (NeurIPS)

    2023

  33. [41]

    Apurva Pathak, Kshitiz Gupta, and Julian McAuley. 2017. Generating and personalizing bundle recommendations on steam. In Proceedings of the 40th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR) . 1073–1076

    2017

  34. [42]

    Yuyang Ren, Zhang Haonan, Luoyi Fu, Xinbing Wang, and Chenghu Zhou. 2023. Distillation-enhanced graph masked autoencoders for bundle recommendation. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR) . 1660–1669

    2023

  35. [43]

    Steffen Rendle, Christoph Freudenthaler, Zeno Gantner, and Lars Schmidt-Thieme. 2009. BPR: Bayesian personalized ranking from implicit feedback. In Proceedings of the Twenty-Fifth Conference on Uncertainty in Artificial Intelligence (UAI) . 452–461

    2009

  36. [44]

    Alireza Salemi, Surya Kallumadi, and Hamed Zamani. 2024. Optimization methods for personalizing large language models through retrieval augmentation. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR) . 75...

    2024

  37. [45]

    Oren Sar Shalom, Noam Koenigstein, Ulrich Paquet, and Hastagiri P Vanchinathan. 2016. Beyond collaborative filtering: The list recommendation problem. In Proceedings of the 25th International Conference on World Wide Web (WWW)) . 63–72

    2016

  38. [46]

    Wenqi Sun, Ruobing Xie, Junjie Zhang, Wayne Xin Zhao, Leyu Lin, and Ji-Rong Wen. 2024. Distillation is all you need for practically using different pre-trained recommendation models. CoRR (2024)

    2024

  39. [47]

    Zhu Sun, Kaidong Feng, Jie Yang, Hui Fang, Xinghua Qu, Yew-Soon Ong, and Wenyuan Liu. 2024. Revisiting bundle recommendation for intent-aware product bundling. ACM Transactions on Recommender Systems (TORS) (2024)

    2024

  40. [48]

    Zhu Sun, Kaidong Feng, Jie Yang, Xinghua Qu, Hui Fang, Yew-Soon Ong, and Wenyuan Liu. 2024. Adaptive in-context learning with large language models for bundle generation. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information R...

    2024

  41. [49]

    Zhu Sun, Hongyang Liu, Xinghua Qu, Kaidong Feng, Yan Wang, and Yew Soon Ong. 2024. Large language models for intent-driven session recommendations. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR) . 324–334

    2024

  42. [50]

    Zhu Sun, Jie Yang, Kaidong Feng, Hui Fang, Xinghua Qu, and Yew Soon Ong. 2022. Revisiting bundle recommendation: Datasets, tasks, challenges and opportunities for intent-aware product bundling. In Proceedings of the 45th International ACM SIGIR Conference on Research and Devel...

    2022

  43. [51]

    Wenjie Wang, Xinyu Lin, Fuli Feng, Xiangnan He, and Tat-Seng Chua. 2023. Generative recommendation: Towards next-generation recommender paradigm. arXiv preprint arXiv:2304.03516 (2023)

    2023 arXiv

  44. [52]

    Xinfeng Wang, Jin Cui, Yoshimi Suzuki, and Fumiyo Fukumoto. 2024. RDRec: Rationale distillation for LLM-based recommendation. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL) . 65–74

    2024

  45. [53]

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2023. Self-consistency improves chain of thought reasoning in language models. In Proceedings of the 11th International Conference on Learning Representations (ICLR)

    2023

  46. [54]

    Yan Wang, Zhixuan Chu, Xin Ouyang, Simeng Wang, Hongyan Hao, Yue Shen, Jinjie Gu, Siqiao Xue, James Y Zhang, Qing Cui, et al. 2023. Enhancing recommender systems with large language model reasoning graphs. arXiv preprint arXiv:2308.10835 (2023)

    2023 arXiv

  47. [55]

    Yuling Wang, Changxin Tian, Binbin Hu, Yanhua Yu, Ziqi Liu, Zhiqiang Zhang, Jun Zhou, Liang Pang, and Xiao Wang. 2024. Can small language models be good reasoners for sequential recommendation?. In Proceedings of the ACM Web Conference 2024 (WWW)

    2024

  48. [56]

    Penghui Wei, Shaoguo Liu, Xuanhua Yang, Liang Wang, and Bo Zheng. 2022. Towards personalized bundle creative generation with contrastive non-autoregressive decoding. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrie...

    2022

  49. [57]

    Yinwei Wei, Xiaohao Liu, Yunshan Ma, Xiang Wang, Liqiang Nie, and Tat-Seng Chua. 2023. Strategy-aware bundle recommender system. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR) . 1198–1207

    2023

  50. [58]

    Chuhan Wu, Fangzhao Wu, Yang Yu, Tao Qi, Yongfeng Huang, and Qi Liu. 2021. NewsBERT: Distilling pre-trained language model for intelligent news application. In Findings of the Association for Computational Linguistics: EMNLP 2021

    2021

  51. [59]

    Likang Wu, Zhaopeng Qiu, Zhi Zheng, Hengshu Zhu, and Enhong Chen. 2024. Exploring large language model for graph data understanding in online job recommendations. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) , Vol. 38. 9178–9186

    2024

  52. [60]

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

    2023 arXiv

  53. [61]

    Min Xie, Laks VS Lakshmanan, and Peter T Wood. 2010. Breaking out of the box of recommendations: From items to packages. In Proceedings of the 4th ACM Conference on Recommender Systems (RecSys) . 151–158

    2010

  54. [62]

    Min Xie, Laks VS Lakshmanan, and Peter T Wood. 2014. Generating top-k packages via preference elicitation. Proceedings of the VLDB Endowment (VLDB) 7, 14 (2014), 1941–1952

    2014

  55. [63]

    Yiyan Xu, Wenjie Wang, Fuli Feng, Yunshan Ma, Jizhi Zhang, and Xiangnan He. 2024. Diffusion models for generative outfit recommendation. In Proceedings of the 47th international ACM SIGIR conference on research and development in information retrieval (SIGIR) . 1350–1359

    2024

  56. [64]

    De-Nian Yang, Wang-Chien Lee, Nai-Hui Chia, Mao Ye, and Hui-Ju Hung. 2012. On bundle configuration for viral marketing in social networks. In Proceedings of the 21st ACM International Conference on Information and Knowledge Management (CIKM) . 2234–2238

    2012

  57. [65]

    Yang Yu, Fangzhao Wu, Chuhan Wu, Jingwei Yi, and Qi Liu. 2021. Tiny-newsrec: Effective and efficient plm-based news recommendation. arXiv preprint arXiv:2112.00944 (2021)

    2021 arXiv

  58. [66]

    Xu Yuan, Hongshen Chen, Yonghao Song, Xiaofang Zhao, Zhuoye Ding, Zhen He, and Bo Long. 2021. Improving sequential recommendation consistency with self-supervised imitation. arXiv preprint arXiv:2106.14031 (2021)

    2021 arXiv

  59. [67]

    Jianyang Zhai, Xiawu Zheng, Chang-Dong Wang, Hui Li, and Yonghong Tian. 2023. Knowledge prompt-tuning for sequential recommendation. In Proceedings of the 29th ACM International Conference on Multimedia (MM) . 6451–6461

    2023

  60. [68]

    Jizhi Zhang, Keqin Bao, Yang Zhang, Wenjie Wang, Fuli Feng, and Xiangnan He. 2023. Is chatgpt fair for recommendation? Evaluating fairness in large language model recommendation. In Proceedings of the 17th ACM Conference on Recommender Systems (RecSys) . 993–999

    2023

  61. [69]

    Sen Zhao, Wei Wei, Ding Zou, and Xianling Mao. 2022. Multi-view intent disentangle graph networks for bundle recommendation. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) . 4379–4387

    2022

  62. [70]

    Tao Zhu, Patrick Harrington, Junjun Li, and Lei Tang. 2014. Bundle recommendation in ecommerce. In Proceedings of the 37th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR) . 657–666. Manuscript submitted to ACM 28 Kaidong Feng, Zh...

    2014

  63. [71]

    Ding Zou, Sen Zhao, Wei Wei, Xian-ling Mao, Ruixuan Li, Dangyang Chen, Rui Fang, and Yuanyuan Fu. 2023. Towards hierarchical intent disentanglement for bundle recommendation. IEEE Transactions on Knowledge and Data Engineering (TKDE) 36, 7 (2023), 3556–3567. Manuscript submitt...

    2023

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.