Pith. sign in

REVIEW 4 major objections 7 minor 1 cited by

SessionRec: Next Session Prediction Paradigm For Generative Sequential Recommendation

T0 review · 4 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Generative recommenders should predict the full set of items a user will engage with in the next session, not the next single item; the paper shows this session-level objective with a rank loss outperforms next-item training on benchmarks…

desk verdict A useful session-level training recipe that improves generative recommenders, but the 'paradigm' claim outruns the ablations. read the letter →

arxiv 2502.10157 v2 pith:JLH3T44X submitted 2025-02-14 cs.IR cs.AI

Lei Huang , Hao Guo , Linzhi Peng , Long Zhang , Xiaoteng Wang , Daoyuan Wang , Shichao Wang , Jinpeng Wang
show 2 more authors
Lei Wang Sheng Chen
This is my paper · ORCID
classification cs.IRcs.AI
keywords generativerecommendationsequentialnextsessionpredictionaggregationranklossscalinglawsretrievalandrankingrecommendersystems
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 argues that generative recommender systems are trained on the wrong prediction target. Instead of predicting the next item a user clicks, as current autoregressive next-item models do, recommenders should predict the full set of items the user will engage with in the next session, because a real recommendation request returns many items at once and user interests are diverse. The paper presents SessionRec, a training paradigm that aggregates items within each session into session-level tokens, encodes the session sequence, and trains with a retrieval-style loss over all positive items in the next session plus a rank loss that pushes positive items above in-session exposures. On two public datasets the paper reports that SessionRec with a transformer-style backbone improves Recall@500 by roughly 35 to 50 percent over the best baselines under an evaluation that holds out the final session, and an online deployment improves payment-related business metrics. If this holds, the next-session objective offers a model-agnostic and computationally cheaper way to build industrial generative recommenders, and it may let one model handle retrieval and ranking together.

What carries the argument

The central machinery is hierarchical session aggregation plus a two-term session-level loss. The Item-based Session Encoder pools item embeddings within each session to form one session token, and the Session-based Sequence Encoder, a decoder-only transformer-style backbone such as HSTU, reads the sequence of session tokens and emits a user interest representation. Scores are dot products between that representation and item embeddings. Training minimizes the retrieval loss, a sampled cross-entropy over all positive items of the next session with random negatives, plus a weighted rank loss that contrasts the same positives against the in-session exposure-only negatives. Session tokenization is what shortens the encoder input by the average session length $M$ and is what makes large numbers of exposure negatives tractable.

What would settle it

On a held-out fold, replace every exposure-only negative in the rank loss with randomly sampled items while keeping everything else fixed; if Recall@500 and NDCG@500 do not degrade, the exposure signal is not doing the work the paper attributes to it.

Watch

Extended reading notes

Core claim

The central claim is that the Next Session Prediction Paradigm, predicting every positively interacted item in the user's next session from a session-aggregated history, is the right generative objective for sequential recommendation and should replace the Next Item Prediction Paradigm. The paper claims that session-level aggregation, implemented as an item-based session encoder followed by a session-based sequence encoder, captures a user's multiple interests at once, cuts attention cost by roughly $M^2$ where $M$ is the average session length, and makes negative exposure interactions cheap to include. Adding a rank loss that contrasts positive items against exposure-only interactions in the next session is reported to sharpen ranking without sacrificing retrieval. The paradigm is claimed to be model-agnostic, improving GRU, Transformer, and HSTU backbones alike, and the paper reports power-law scaling with training data and online A/B gains from deploying SessionRec in the retrieval stage of a large consumer app.

Load-bearing premise

The results assume that exposure-only interactions mean the user was not interested, even though those exposures were chosen by a previous recommender and are subject to position and selection bias.

Editorial extensions

If this is right

  • Training objectives for generative recommenders should be session-level: every positive item in the next session is a label, not merely the next item.
  • Because session aggregation shrinks the encoder input by the average session length, transformer-style backbones become much cheaper, making full exposure histories feasible in training and serving.
  • The rank loss over in-session exposures gives a single generative model a ranking capability, pointing toward replacing cascaded retrieval and ranking systems with one model.
  • The reported power-law scaling implies session-level generative recommenders continue to improve as training data and compute grow.
  • The model-agnostic formulation means existing backbones can be upgraded by changing the prediction target and adding session aggregation rather than redesigning the architecture.

Reading between the lines

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

  • Editorial inference: because exposure-only negatives were generated by an older recommender's policy, part of the reported gain could come from fitting that policy rather than from true user disinterest; randomized exposure logging would separate the two.
  • Editorial inference: the same session-level tokenization could transfer to other domains where outputs arrive in batches, such as feed ranking or notification selection.
  • Editorial inference: since simple mean pooling matched or beat more complex session encoders, the gains may come largely from the prediction objective rather than from the aggregation network; an ablation keeping item-level tokens while using the session-level loss would test this.
  • Editorial inference: the final-session held-out evaluation the paper adopts is a more realistic benchmark for multi-item recommendation, but very long sessions may inflate recall because more labels are easy to hit.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 7 minor

Summary. The paper introduces SessionRec, a next-session prediction paradigm (NSPP) for generative sequential recommendation. Instead of predicting the next item autoregressively, the model hierarchically aggregates user interactions into session-level representations (Item-based Session Encoder followed by Session-based Sequence Encoder) and is trained with a sampled cross-entropy retrieval loss over all positive items of the next session, plus a rank loss that uses exposure-only interactions in the next session as negatives. Experiments on KuaiSAR and RecFlow report consistent gains over GRU4Rec, SASRec, SASRec+, BERT4Rec, HSTU, and HSTU+ under both leave-one-item-out and leave-one-session-out evaluation, using full unsampled Recall@K and NDCG@K. The paper also reports an online A/B test at Meituan, a scaling-law curve with training data size, and an efficiency comparison. The central claim is that the next-session prediction objective, not simply the hierarchical aggregation or the inclusion of negative interactions, drives the improvements.

Significance. If the central claim is validated, the paper would be a practically relevant contribution: it proposes a model-agnostic training and representation paradigm that can be layered on GRU, Transformer, or HSTU backbones, reduces the sequence length processed by the backbone by the average session length, and reports sizable gains over strong baselines. The use of full unsampled metrics, the inclusion of both next-item and next-session evaluation strategies, the public-dataset reproducibility plan, and the online deployment results are concrete strengths. The main risk is that the reported comparisons do not isolate the next-session prediction objective from the session-level input representation, so the title-level claim that the paradigm shift itself is responsible for the gains is not yet supported.

major comments (4)
  1. [Section 4.2, Tables 2 and 3] The central attribution of the gains to the next-session paradigm is underdetermined because SessionRec-HSTU differs from HSTU+ in at least three ways: hierarchical session aggregation (ISE+SSE), the session-level objective in Eq. (8) and Eq. (9), and the use of exposure-only interactions as negatives in the rank loss. No experiment holds the architecture fixed and varies only the prediction objective. In particular, a model with the same ISE+SSE architecture but a next-item objective, and a model with item-level encoding but the Eq. (8) session objective, are both missing. Without these two cells, the observed improvements in Tables 2 and 3 could be driven largely by the aggregation and its induced representational/computational regularization rather than by next-session prediction. Please add the missing 2x2 ablation, at least on KuaiSAR, and report how each component contributes to the headline gains.
  2. [Equation (9), Section 4.1.1, Section 4.5] The rank loss treats all exposure-only interactions as negatives, but exposure behavior is generated by the logged recommender policy and is subject to position bias, selection bias, and the policy's own retrieval distribution; it is not a random sample of user disinterest. Since the paper credits the rank loss with the ranking improvements in Figure 4, this assumption is load-bearing for the ranking claim. Please provide an analysis or a robustness check, for example by reweighting exposure negatives, downsampling them, or comparing against a version that uses only random negatives. At minimum, state explicitly that the model inherits the logged policy's biases and discuss how they affect the reported gains.
  3. [Tables 2 and 3, Section 6] No repeated runs, confidence intervals, or significance tests are reported for any offline metric, and the online A/B section reports only point estimates (+0.603% Pay PV, +0.564% PVCTCVR) without p-values, confidence intervals, or the number of users/traffic split. Given that the largest headline offline gain is reported as Recall@500 and the online gains are around 0.5%, the reader cannot assess whether the differences are stable or within run-to-run and day-to-day variance. Please report variance across seeds for the main comparisons and standard error or confidence intervals for the online metrics.
  4. [Section 5, Figure 5] The scaling-law claim is supported only by a fitted curve over a small number of observed points. The fitted log-linear curve is computed from the same data points that are then presented as evidence of the scaling law, and no holdout data, error bars, or comparison with alternative functional forms are given. The statement that 'the performance of the model shows a trend of linear increase' with exponential data growth is a reasonable observation from the plotted points, but the paper should not present this as an established power-law scaling law. Please either reframe the claim as an empirical observation on one configuration or provide a more rigorous study with multiple seeds and held-out data sizes.
minor comments (7)
  1. [Section 4.1.1] The dataset statistics in Table 1 report 9,460 users for KuaiSAR, while the text says the raw dataset contains interactions of 25,877 users; please clarify that the table reports post-filtering statistics.
  2. [Section 4.1.4] The implementation details state that 'all other methods incorporate 128 negative samples per item' but the negative-sampling procedure for baselines such as SASRec+ is not fully specified; please clarify whether negatives are sampled per target and whether the same count is used for SessionRec.
  3. [Abstract and Section 1] The abstract and introduction claim an average 27% improvement over the best baseline, but the percentages in Table 2 vary widely by metric and dataset; please state which metric set the average is computed over.
  4. [Section 3.6] There is a typo in 'DotProduct represents the commonly useded operation'; please correct it.
  5. [Section 2 and Table 2] The baseline name is written as 'Bert4Rec' in Table 2 and the text but 'BERT4Rec' elsewhere; please unify the capitalization.
  6. [Section 4.1.2] The description of SASRec+ says 'an improved version of SAS4Rec'; please use the correct model name 'SASRec'.
  7. [Section 4.2, Figure 3] Figure 3 mixes training time and Recall@500 on two axes; please state explicitly which dataset this figure is for and add units for the horizontal axis.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claims are empirical, held-out evaluations rather than derivations from the method's own outputs.

full rationale

SessionRec's main claims, that next-session prediction improves recall and NDCG on KuaiSAR and RecFlow and yields online gains at Meituan, are supported by supervised training on logged sessions and evaluation on held-out sessions under leave-one-session-out and leave-one-item-out protocols. The training objective (Eq. 8) and rank loss (Eq. 9) use labels from the next session, and the metrics (Recall@K, NDCG@K) measure the same target, but this is standard supervised learning, not circularity: the model must generalize to unseen sessions, and baselines are evaluated under identical protocols. The reported improvements are empirical measurements, not consequences derivable from the equations alone. The scaling-law plot (Figure 5) fits a log curve to four observed data-size points; it is a descriptive curve fit, not a prediction forced by construction, and it is peripheral to the main claims. The only self-citation (Ref. [3], to a SimHash method by co-author Sheng Chen et al.) appears in a related-work list and is not load-bearing. Hyperparameter choices such as the rank-loss weight alpha are selected by sweeping on the same datasets used for reporting, which is a model-selection concern, but selecting alpha does not make the reported test-set metrics equal to the training objective by construction. The absence of a fully crossed architecture/objective control is an underdetermination concern, not circularity. No load-bearing step reduces to its own inputs, and no uniqueness claim or ansatz is imported from the authors' prior work.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central empirical claims rest on a standard encoder-predictor architecture, with no new theoretical entities. The load-bearing assumptions are about the meaning of session IDs and exposure-based negatives, the independence of positive targets in the retrieval loss, and the adequacy of a single dot-product user embedding. Several hyperparameters, most notably the rank-loss weight alpha, are selected on the validation data, and the claimed scaling law is a curve fit to five points.

free parameters (5)
  • Rank loss weight alpha = 0.2 (best on validation; grid 0, 0.05, 0.1, 0.2, 0.5, 1, 2)
    Chosen per dataset to balance retrieval and ranking in Eq. (10); reported improvements depend on this choice.
  • Negative sampling size C = 128
    Used in retrieval loss Eq. (8) for all methods except SASRec; affects the difficulty of the sampled softmax.
  • Session encoder aggregation = mean pooling
    Selected after Table 4 comparisons; the paradigm's efficiency and accuracy depend on this choice.
  • Max behavior sequence length = 200 (KuaiSAR), 400 (RecFlow), 8000 (online)
    Truncation choices for input sequences in offline and online settings.
  • Scaling-law fit coefficients = a=0.03829156, b=-0.348426
    Fitted to five data points in Figure 5; the scaling-law claim is a curve fit, not a parameter-free prediction.
assumptions (5)
  • domain assumption Exposure-only interactions in the logged data can be treated as true negative preferences for the rank loss.
    Section 4.1.1 defines negative interactions as exposure-only; Section 3.7.2 uses them as hard negatives in Eq. (9). Logged exposure depends on the previous recommender's policy and position, so it may not be a clean negative signal.
  • domain assumption Positive items in the next session are independent targets under a per-item cross-entropy retrieval loss.
    Eq. (8) sums per-positive-item softmax losses with random negatives; no joint distribution over the set of next-session items is modeled. If items within a session are correlated, the objective is a surrogate rather than a likelihood.
  • domain assumption Session identifiers in the data define meaningful prediction units and are available at serving time.
    The problem statement in Section 3.1 assumes session IDs; preprocessing filters to sessions with at least one positive and users with at least three sessions. Noisy or absent session boundaries would break the aggregation.
  • domain assumption Dot-product similarity between a single user embedding and item embeddings is sufficient to rank all candidates for the next session.
    Eq. (7) and the deployment in Section 6 score every item by a dot product between user and item embeddings; this two-tower assumption may limit expressiveness for diverse multi-interest sessions.
  • domain assumption Causal decoder-only SSE prevents temporal leakage across sessions.
    Section 3.5 advises decoder-only SSE to prevent temporal leakage; the masking is standard, but the paper does not show a leakage audit or compare with bidirectional encoding.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SessionRec: Next Session Prediction Paradigm For Generative Sequential Recommendation." pith.science (2026). https://pith.science/paper/JLH3T44X

@misc{pith2026250210157,
  author       = {Pith},
  title        = {Pith review of: SessionRec: Next Session Prediction Paradigm For Generative Sequential Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JLH3T44X}},
  note         = {Machine review of arXiv:2502.10157}
}
read the original abstract

We introduce SessionRec, a novel next-session prediction paradigm (NSPP) for generative sequential recommendation, addressing the fundamental misalignment between conventional next-item prediction paradigm (NIPP) and real-world recommendation scenarios. Unlike NIPP's item-level autoregressive generation that contradicts actual session-based user interactions, our framework introduces a session-aware representation learning through hierarchical sequence aggregation (intra/inter-session), reducing attention computation complexity while enabling implicit modeling of massive negative interactions, and a session-based prediction objective that better captures users' diverse interests through multi-item recommendation in next sessions. Moreover, we found that incorporating a rank loss for items within the session under the next session prediction paradigm can significantly improve the ranking effectiveness of generative sequence recommendation models. We also verified that SessionRec exhibits clear power-law scaling laws similar to those observed in LLMs. Extensive experiments conducted on public datasets and online A/B test in Meituan App demonstrate the effectiveness of SessionRec. The proposed paradigm establishes new foundations for developing industrial-scale generative recommendation systems through its model-agnostic architecture and computational efficiency.

Figures

Figures reproduced from arXiv: 2502.10157 by the authors.

Figure 1
Figure 1. Next Item Prediction Paradigm (NIPP) vs. our pro [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overall architecture of our proposed SessionRec. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Performance and Traning Time per Epoch SessionRec-HSTU predicts diverse interests across a broader space, which enables simultaneous enhancement of prediction accuracy at both session and item levels. HSTU-based models demonstrate the best performance, followed by Transformer-based models, with GRU-based models performing the worst. Increasing the complexity of the model can enhance the accuracy of item-level models… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Impact of ranking loss weight on model perfor [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 6
Figure 6. Figure 6: Deployment architecture of SessionRec in online [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. PinFM: Foundation Model for User Activity Sequences at a Billion-scale Visual Discovery Platform

    cs.LG 2025-07 conditional novelty 5.0 of 10

    A single pretrained model over user activity sequences improves save rates in Pinterest's Home Feed and Related Items ranking when fine-tuned per application, while deduplication and quantization keep serving costs neutral.

Reference graph

Works this paper leans on

48 extracted references · 8 canonical work pages · cited by 1 Pith paper

  1. [1]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Floren- cia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)

    arXiv 2023

  2. [2]

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems 33 (2020), 1877–1901

    2020

  3. [3]

    Yue Cao, Xiaojiang Zhou, Jiaqi Feng, Peihao Huang, Yao Xiao, Dayao Chen, and Sheng Chen. 2022. Sampling is all you need on modeling long-term user behaviors for CTR prediction. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management . 2974–2983

    2022

  4. [4]

    Jianxin Chang, Chenbin Zhang, Zhiyi Fu, Xiaoxue Zang, Lin Guan, Jing Lu, Yiqun Hui, Dewei Leng, Yanan Niu, Yang Song, et al. 2023. TWIN: TWo-stage interest network for lifelong user behavior modeling in CTR prediction at kuaishou. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 3785–3794

    2023

  5. [5]

    Junyi Chen, Lu Chi, Bingyue Peng, and Zehuan Yuan. 2024. Hllm: Enhancing sequential recommendations via hierarchical large language models for item and user modeling. arXiv preprint arXiv:2409.12740 (2024)

    arXiv 2024

  6. [6]

    Qiwei Chen, Changhua Pei, Shanshan Lv, Chao Li, Junfeng Ge, and Wenwu Ou

  7. [7]

    Heng-Tze Cheng, Levent Koc, Jeremiah Harmsen, Tal Shaked, Tushar Chandra, Hrishi Aradhye, Glen Anderson, Greg Corrado, Wei Chai, Mustafa Ispir, et al

  8. [8]

    Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. 2014. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555 (2014)

    arXiv 2014

Show all 48 references
  1. [9]

    Sumanth Doddapaneni, Krishna Sayana, Ambarish Jash, Sukhdeep Sodhi, and Dima Kuzmin. 2024. User Embedding Model for Personalized Language Prompt- ing. arXiv preprint arXiv:2401.04858 (2024)

    2024 arXiv

  2. [10]

    Alexey Dosovitskiy. 2020. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)

    2020 arXiv

  3. [11]

    Yufei Feng, Fuyu Lv, Weichen Shen, Menghan Wang, Fei Sun, Yu Zhu, and Keping Yang. 2019. Deep session interest network for click-through rate prediction. arXiv preprint arXiv:1905.06482 (2019)

    2019 arXiv

  4. [12]

    F Maxwell Harper and Joseph A Konstan. 2015. The movielens datasets: History and context. Acm transactions on interactive intelligent systems (tiis) 5, 4 (2015), 1–19

    2015

  5. [13]

    Balázs Hidasi and Alexandros Karatzoglou. 2018. Recurrent neural networks with top-k gains for session-based recommendations. In Proceedings of the 27th ACM international conference on information and knowledge management . 843–852

    2018

  6. [14]

    Jeff Johnson, Matthijs Douze, and Hervé Jégou. 2019. Billion-scale similarity search with GPUs. IEEE Transactions on Big Data 7, 3 (2019), 535–547

    2019

  7. [15]

    Wang-Cheng Kang and Julian McAuley. 2018. Self-attentive sequential recom- mendation. In 2018 IEEE international conference on data mining (ICDM) . IEEE, 197–206

    2018

  8. [16]

    Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. 2020. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361 (2020)

    2020 arXiv

  9. [17]

    Anton Klenitskiy and Alexey Vasilev. 2023. Turning Dross Into Gold Loss: is BERT4Rec really better than SASRec?. In Proceedings of the 17th ACM Conference on Recommender Systems. 1120–1125

    2023

  10. [18]

    Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. 2022. Autoregressive image generation using residual quantization. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 11523–11532

    2022

  11. [19]

    Jianxun Lian, Xiaohuan Zhou, Fuzheng Zhang, Zhongxia Chen, Xing Xie, and Guangzhong Sun. 2018. xdeepfm: Combining explicit and implicit feature in- teractions for recommender systems. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data ...

    2018

  12. [20]

    Bing Liu, Wynne Hsu, and Yiming Ma. 1998. Integrating classification and association rule mining. In Proceedings of the fourth international conference on knowledge discovery and data mining . 80–86

    1998

  13. [21]

    Qi Liu, Kai Zheng, Rui Huang, Wuchao Li, Kuo Cai, Yuan Chai, Yanan Niu, Yiqun Hui, Bing Han, Na Mou, et al . 2024. RecFlow: An Industrial Full Flow Recommendation Dataset. arXiv preprint arXiv:2410.20868 (2024)

    2024 arXiv

  14. [22]

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. 2021. Swin transformer: Hierarchical vision transformer us- ing shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision. 10012–10022

    2021

  15. [23]

    Yu A Malkov and Dmitry A Yashunin. 2018. Efficient and robust approximate nearest neighbor search using hierarchical navigable small world graphs. IEEE transactions on pattern analysis and machine intelligence 42, 4 (2018), 824–836

    2018

  16. [24]

    Julian McAuley, Christopher Targett, Qinfeng Shi, and Anton Van Den Hengel

  17. [25]

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback. Advances in neural information processing systems 35...

    2022

  18. [26]

    Qi Pi, Weijie Bian, Guorui Zhou, Xiaoqiang Zhu, and Kun Gai. 2019. Practice on long sequential user behavior modeling for click-through rate prediction. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining . 2671–2679

    2019

  19. [27]

    Qi Pi, Guorui Zhou, Yujing Zhang, Zhe Wang, Lejian Ren, Ying Fan, Xiaoqiang Zhu, and Kun Gai. 2020. Search-based user interest modeling with lifelong sequential behavior data for click-through rate prediction. In Proceedings of the 29th ACM International Conference on Informat...

    2020

  20. [28]

    Alec Radford. 2018. Improving language understanding by generative pre- training. (2018)

    2018

  21. [29]

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI blog 1, 8 (2019), 9

    2019

  22. [30]

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

  23. [31]

    Weiping Song, Chence Shi, Zhiping Xiao, Zhijian Duan, Yewen Xu, Ming Zhang, and Jian Tang. 2019. Autoint: Automatic feature interaction learning via self- attentive neural networks. InProceedings of the 28th ACM international conference on information and knowledge management ...

    2019

  24. [32]

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

  25. [33]

    Jiaxi Tang and Ke Wang. 2018. Personalized top-n sequential recommenda- tion via convolutional sequence embedding. In Proceedings of the eleventh ACM international conference on web search and data mining . 565–573

    2018

  26. [34]

    A Waswani, N Shazeer, N Parmar, J Uszkoreit, L Jones, A Gomez, L Kaiser, and I Polosukhin. 2017. Attention is all you need. In NIPS

    2017

  27. [35]

    Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. 2021. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652 (2021)

    2021 arXiv

  28. [36]

    Chao-Yuan Wu, Amr Ahmed, Alex Beutel, Alexander J Smola, and How Jing. 2017. Recurrent recommender networks. In Proceedings of the tenth ACM international conference on web search and data mining . 495–503

    2017

  29. [37]

    Li Yang, Anushya Subbiah, Hardik Patel, Judith Yue Li, Yanwei Song, Reza Mirghaderi, and Vikram Aggarwal. 2024. Item-Language Model for Conver- sational Recommendation. arXiv preprint arXiv:2406.02844 (2024)

    2024 arXiv

  30. [38]

    Yong Yu, Xiaosheng Si, Changhua Hu, and Jianxun Zhang. 2019. A review of recurrent neural networks: LSTM cells and network architectures. Neural computation 31, 7 (2019), 1235–1270

    2019

  31. [39]

    Jiaqi Zhai, Zhaojie Gong, Yueming Wang, Xiao Sun, Zheng Yan, Fu Li, and Xing Liu. 2023. Revisiting Neural Retrieval on Accelerators. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 5520–5531

    2023

  32. [40]

    Jiaqi Zhai, Lucy Liao, Xing Liu, Yueming Wang, Rui Li, Xuan Cao, Leon Gao, Zhaojie Gong, Fangda Gu, Jiayuan He, Yinghai Lu, and Yu Shi. 2024. Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations. In Proceedings of the 41st I...

    2024

  33. [41]

    Tian Zhang, Raghu Ramakrishnan, and Miron Livny. 1996. BIRCH: an efficient data clustering method for very large databases. ACM sigmod record 25, 2 (1996), 103–114

    1996

  34. [42]

    Guorui Zhou, Na Mou, Ying Fan, Qi Pi, Weijie Bian, Chang Zhou, Xiaoqiang Zhu, and Kun Gai. 2019. Deep interest evolution network for click-through rate prediction. In Proceedings of the AAAI conference on artificial intelligence , Vol. 33. 5941–5948

    2019

  35. [43]

    Guorui Zhou, Xiaoqiang Zhu, Chenru Song, Ying Fan, Han Zhu, Xiao Ma, Yanghui Yan, Junqi Jin, Han Li, and Kun Gai. 2018. Deep interest network for click-through rate prediction. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining ...

    2018

  36. [2015]

    In Proceedings of the 38th international ACM SIGIR conference on research and development in information retrieval

    Image-based recommendations on styles and substitutes. In Proceedings of the 38th international ACM SIGIR conference on research and development in information retrieval. 43–52

  37. [2016]

    In Proceedings of the 1st workshop on deep learning for recommender systems

    Wide & deep learning for recommender systems. In Proceedings of the 1st workshop on deep learning for recommender systems . 7–10

  38. [2019]

    In Proceedings of the 28th ACM international conference on information and knowledge management

    BERT4Rec: Sequential recommendation with bidirectional encoder rep- resentations from transformer. In Proceedings of the 28th ACM international conference on information and knowledge management . 1441–1450

  39. [2021]

    arXiv preprint arXiv:2108.04468 (2021)

    End-to-end user behavior retrieval in click-through rateprediction model. arXiv preprint arXiv:2108.04468 (2021)

    2021 arXiv

  40. [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

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.