Pith. sign in

REVIEW 2 major objections 6 minor 1 cited by

Action is All You Need: Dual-Flow Generative Ranking Network for Recommendation

T0 review · 2 major / 6 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 Two-flow attention cuts generative ranking cost while improving accuracy.

desk verdict Dual-flow training for HSTU ranking is a genuine architectural contribution and the efficiency math holds; the accuracy claim over MetaGR is unproven without error bars, and the train/serve gap is unexamined. read the letter →

arxiv 2505.16752 v3 pith:WOHXBW2W submitted 2025-05-22 cs.IR cs.AI

Hao Guo , Erpeng Xue , Lei Huang , Shichao Wang , Xiaolei Wang , Lei Wang , Jinpeng Wang , Sheng Chen This is my paper
classification cs.IRcs.AI
keywords dual-flowarchitecturegenerativerankingrecommendationsystemself-attentionHSTUCTRpredictiontrainingefficiencyscalinglaw
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

Generative ranking models learn directly from raw user behavior sequences, but the interleaved token layout used by current HSTU-based baselines turns each interaction into two tokens, doubling sequence length and inflating the quadratic attention cost. This paper's central proposal, the Dual-Flow Generative Ranking Network (DFGR), duplicates the sequence into a real flow carrying true action types and a fake flow carrying placeholder actions, sharing weights between the two. During training the fake flow attends with its own queries but borrows keys and values from the real flow for all preceding positions, so each position sees full context without seeing the label. That arrangement lets one interaction occupy one token, cutting training compute to roughly half and inference compute to about a quarter of the interleaved baseline while reporting better AUC and G-AUC on public and industrial datasets.

What carries the argument

The load-bearing object is the cross-flow KV borrowing in the QKV module of self-attention. In the fake flow, the query at position $i$ is projected normally, but the keys and values attending over positions $1...i-1$ are taken from the real flow's hidden states, so the placeholder token receives the same contextual information a true action token would, without being told the answer. The current position uses the fake flow's own key and value, preserving the causal mask. This single mechanism simultaneously prevents label leakage, halves the token sequence length relative to interleaved item/action encoding, and enables the same merged single-flow forward pass at serving time. A session-aware cross-triangle mask zeroes attention among items sharing a session so candidates can be scored in parallel without mutual information.

What would settle it

Train the same architecture with the loss computed on real-flow outputs instead of fake-flow outputs while keeping the single-flow inference path, or run a live A/B test comparing the dual-flow-trained model against a model trained directly on the serving-time input format; if the reported AUC/G-AUC advantage disappears, the central transfer assumption fails.

Watch

Extended reading notes

Core claim

DFGR's claim is that separating 'what the user did' from 'what the model may observe' into two parallel attention streams is strictly more efficient and at least as accurate as encoding each action as a separate token. In training, the real flow performs standard causal self-attention over items enriched with true action types; the fake flow computes a standard forward pass of its own but, for positions before the current one, replaces its keys and values with those from the real flow's corresponding layer, using its own keys and values only at the current position. The loss is computed only on fake-flow outputs, so the model learns to predict the next action while the real flow supplies complete historical context. At inference, the same merged single-flow sequence used by the single-flow variant is scored in one forward pass, with attention zeroed out among candidate items in the same session. The reported result is that DFGR beats the interleaved-token HSTU baseline by 0.31%--1.2% AUC and 0.11%--1.4% G-AUC across three datasets, and beats the feature-engineered industrial DLRM on the TRec dataset (AUC 0.8755 vs. 0.8703), while reducing theoretical training complexity from $4O(N^2)$ to $2O(N^2)$ and inference complexity from $4O(N^2)$ to $O(N^2)$.

Load-bearing premise

The dual-flow training procedure is assumed to transfer to the single-flow inference procedure; if attending with borrowed real-flow keys and values during training is not equivalent to attending with masked placeholders during serving, the reported offline gains may not survive under the actual serving path.

Editorial extensions

If this is right

  • DFGR cuts theoretical training cost from $4O(N^2)$ to $2O(N^2)$ and inference cost from $4O(N^2)$ to $O(N^2)$ relative to the interleaved-token HSTU baseline, so longer user histories become affordable in both stages.
  • Because each interaction is one token, the same decoder-only backbone can score thousands of candidates in a single forward pass by concatenating them to the user history and masking intra-session attention.
  • Under fixed compute budgets, the optimal depth grows as compute grows; the paper observes the best G-AUC shifting from 2-layer to 4-layer to 8-layer networks as PetaFLOPS increase.
  • G-AUC improves roughly linearly in the logarithm of compute in the plotted range, i.e., a scaling-law-like trend appears for this ranking model.
  • Full-domain behavior data and a Mixture-of-Attention split of shared and channel-specific heads add 0.57% AUC / 0.84% G-AUC and 0.28% G-AUC respectively on the industrial dataset.

Reading between the lines

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

  • The real/fake flow trick is a general way to feed side information through self-attention without letting it leak; it could apply to any sequential recommender or sequence model where categorical labels appear in the input stream, not just ranking.
  • The paper does not analyze the train/serve distribution shift between dual-flow training and single-flow inference; testing that equivalence directly would determine whether the offline gains hold online.
  • The complexity saving depends on the quadratic self-attention term dominating; for very long sequences where linear-attention or retrieval shortcuts would be used, the ratio between approaches could change.
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

2 major / 6 minor

Summary. The paper introduces DFGR, a dual-flow generative ranking architecture for recommendation. DFGR duplicates the user behavior sequence into a real flow (with true action types) and a fake flow (with placeholder action types) that share parameters; during training, the fake flow borrows keys and values from the real flow for all preceding positions, while at inference the model falls back to a single-flow (SFGR) decoder that processes history with real actions and candidates with masked placeholders. The authors claim that DFGR outperforms Meta's HSTU-based MetaGR on AUC/G-AUC while reducing training compute to about half and inference compute to a quarter, and they report experiments on RecFlow, KuaiSAR, and an industrial TRec dataset, together with ablations on full-domain data, mixture-of-attention, learning-rate scheduling, parameter allocation, and a scaling-law analysis.

Significance. If the claims hold, the paper makes a useful efficiency contribution: the complexity analysis in Section 3.3.2 is transparent, is grounded in MetaGR's published formulation, and correctly identifies the quadratic-attention saving from replacing interleaved item/action tokens with two parallel flows. The proposed architecture is simple to describe and could be adopted by other generative ranking systems. The paper also deserves credit for evaluating on both public datasets and an industrial dataset. However, the accuracy advantage over MetaGR is the load-bearing claim, and it currently rests on single unreplicated runs with no error bars or significance tests; the public-dataset G-AUC differences are as small as 0.0008 and 0.0009. The train/serve distribution shift between DFGR training and SFGR inference is also unanalyzed. These issues prevent the empirical superiority claim from being established at the level the paper asserts.

major comments (2)
  1. [Section 4.2, Tables 2 and 3] The central claim that DFGR outperforms MetaGR is supported only by point estimates without variance information. In Table 2, the RecFlow G-AUC difference is 0.6892 vs 0.6884 (0.0008 absolute) and the KuaiSAR G-AUC difference is 0.6081 vs 0.6072 (0.0009 absolute). The paper does not report the number of seeds, standard deviations, confidence intervals, or significance tests, and no code is released. Given that Section 4.2 itself states that 0.1% AUC/G-AUC improvement is considered significant, the paper must demonstrate that these differences are not within run-to-run optimizer or initialization noise. I recommend adding multi-seed results with error bars, reporting paired significance tests, and/or releasing code so the reader can replicate the comparisons. Without this, the headline accuracy gain over MetaGR is not statistically distinguishable from noise.
  2. [Sections 3.3.1 and 3.2.1] There is an unanalyzed train/serve distribution shift. During DFGR training, the fake flow uses placeholder action tokens for every position, including historical positions, and obtains keys/values for all preceding positions from the real flow. During inference, the paper adopts the SFGR strategy in which historical tokens carry real action types and only candidate tokens carry masked placeholders. These are different input distributions: the inference-time history has real actions, whereas the training-time fake flow never sees real actions in its own input stream. The paper does not prove that the two procedures are equivalent, nor does it report an evaluation of the actual serving-consistent variant (for example, fine-tuning or training the model on the exact inference input format). This issue directly affects whether the reported offline gains transfer to the deployed single-flow inference path and should be addressed either by a formal equivalence argument or by serving-consistent experiments.
minor comments (6)
  1. [Section 3.3.2] The sentence 'the ratio of computational complexity between our DFGR and MetaGR is about 2' is inconsistent with the immediately following conclusion that DFGR saves half the training compute. The dominant quadratic terms give a DFGR-to-MetaGR ratio of about 1/2; the intended statement is presumably that MetaGR is about twice as expensive as DFGR. Please correct the wording to avoid ambiguity.
  2. [Section 4.1] There is a typo in 'Datesets' (should be 'Datasets'), and the industrial dataset is referred to as both 'TRec' and 'Trec' inconsistently. Please unify the terminology.
  3. [Related Work] The citation in 'User Behavior Sequence Modeling' contains an unresolved placeholder '[32? ?]'. Please complete the reference to the cited work.
  4. [Section 4.5, Figure 9] The scaling-law claim is based on a log fit to only five measured configurations. The fitted curve has no uncertainty estimates, and the paper does not report points beyond the five depth/width settings. Please either add confidence intervals or soften the claim from 'adheres to the scaling law' to 'is consistent with a log-linear trend on this dataset'.
  5. [General reproducibility] No code is released and the TRec dataset is not publicly available. Please state these limitations explicitly and, if possible, release code for the public-dataset experiments so that the RecFlow and KuaiSAR results can be independently reproduced.
  6. [Section 3.3.1, Eq. (7)] The use of the term (1_t - I_t) to exclude the diagonal cross-flow attention is not explained in the text. Please add a sentence clarifying that this prevents the fake flow from attending to the real flow's action token at the current position, which would leak the label.

Circularity Check

1 steps flagged · score 1.0 of 10

No significant circularity: the efficiency claims are analytic complexity derivations from HSTU's stated FLOPs, and the accuracy claims are empirical comparisons; the only mild self-confirmation is the Figure 9 scaling-law log-fit to the paper's own measured points, which is not load-bearing.

  1. fitted input called prediction [Section 4.5 (RQ4), Figure 9 caption]
    "We plot the computational complexity (represented on a logarithmic scale as the horizontal axis) against the corresponding optimal G-AUC values (vertical axis), as illustrated in Figure 9. The results reveal that in the current scenario, the DFGR model adheres to the scaling law principle, where the G-AUC metric exhibits a linear improvement trend as computational complexity increases following a power-law pattern."

    The asserted scaling law is a post-hoc log-linear fit (caption: 'Fitted Curve: y = 0.0077*log(x) + 0.6774') made through the paper's own measured optimal G-AUC points. The 'linear improvement vs log compute' pattern is therefore a property of the fitting function applied to the same data that is then said to exhibit the law — the finding is partly constructed by the fitting procedure rather than predicted or independently validated. This is a mild self-confirmation only: it does not support the headline accuracy/efficiency claims, which rest on the derivations in Sections 3.2.2/3.3.2 and the empirical tables.

full rationale

The paper's central derivations are self-contained. Table 1's complexity ratios are obtained analytically in Sections 3.2.2 and 3.3.2 from the HSTU FLOPs formula under the stated N >> D approximation: MetaGR's interleaved item/action tokens give length 2N and hence 4·O(N²) attention cost, while DFGR trains two length-N flows (2·O(N²)) and serves one length-N flow (O(N²)); the 4x inference and ~2x training savings follow algebraically and are externally checkable, not fitted. The accuracy comparisons (Tables 2, 3) are empirical runs against external baselines; no parameter is fitted to the reported AUC/G-AUC, so no prediction reduces to its input by construction. The dual-flow mechanism (Eqs. 1-7) is a constructive training design: the fake flow borrows real-flow KV for preceding positions and its own KV for the current position, with loss computed only on the fake flow, and the inference-time single-flow procedure mirrors the same token structure (real actions for history, placeholders for candidates); whether the train/serve procedures are exactly equivalent is an unanalyzed deployment-risk question, not circularity. The only self-citation with overlapping authorship is ref. [1] (Cao et al., CIKM 2022, co-authored by Sheng Chen), which appears solely in a related-work list of DLRM sequence models and is not load-bearing. The sole mild self-confirmation is the Figure 9 scaling-law log-fit described in the step above; it is secondary to the paper's main claims. Statistical caveats (single unreplicated runs, no error bars, metric gaps as small as 0.0008 G-AUC) are correctness and robustness risks, not circularity. Overall the derivation chain does not reduce to its inputs.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The only fitted numbers are the scaling-law curve and manually chosen learning-rate schedule; neither is part of the efficiency derivation. The architecture itself introduces no external postulated entities beyond the dual-flow construct.

free parameters (2)
  • learning rate schedule = constant 5e-4 then decay 5e-6 per 1k steps
    Chosen by hand for DFGR training in Section 4.3.3; affects reported gains and is not derived from first principles.
  • scaling-law log-fit coefficients = y = 0.0077*log(x) + 0.6774
    Fitted to the paper's own measured G-AUC points in Figure 9 to support the scaling-law claim; it is an interpretation of the data, not a prediction.
assumptions (3)
  • standard math Self-attention compute is dominated by the quadratic term and sequence length N is much larger than hidden dimension D.
    Used in Section 3.3.2 to reduce complexity formulas; if N is not much larger than D, the stated 2x/4x ratios degrade.
  • domain assumption Items in one session must be mutually unaware during pointwise scoring.
    Stated in Section 3.3.1 (Cross-Triangle Matrix): zeroing attention within session prevents label leakage; if session-internal dependence is informative, the mask would hurt.
  • domain assumption AUC/G-AUC gains of 0.1% are commercially significant.
    Stated at the start of Section 4.2 and used to frame sub-0.01 absolute AUC differences as meaningful.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Action is All You Need: Dual-Flow Generative Ranking Network for Recommendation." pith.science (2026). https://pith.science/paper/WOHXBW2W

@misc{pith2026250516752,
  author       = {Pith},
  title        = {Pith review of: Action is All You Need: Dual-Flow Generative Ranking Network for Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WOHXBW2W}},
  note         = {Machine review of arXiv:2505.16752}
}
read the original abstract

Deep Learning Recommendation Models (DLRMs) often rely on extensive manual feature engineering to improve accuracy and user experience, which increases system complexity and limits scalability of model performance with respect to computational resources. Recently, Meta introduced a generative ranking paradigm based on HSTU block that enables end-to-end learning from raw user behavior sequences and demonstrates scaling law on large datasets that can be regarded as the state-of-the-art (SOTA). However, splitting user behaviors into interleaved item and action information significantly increases the input sequence length, which adversely affects both training and inference efficiency. To address this issue, we propose the Dual-Flow Generative Ranking Network (DFGR), that employs a dual-flow mechanism to optimize interaction modeling, ensuring efficient training and inference through end-to-end token processing. DFGR duplicates the original user behavior sequence into a real flow and a fake flow based on the authenticity of the action information, and then defines a novel interaction method between the real flow and the fake flow within the QKV module of the self-attention mechanism. This design reduces computational overhead and improves both training efficiency and inference performance compared to Meta's HSTU-based model. Experiments on both open-source and real industrial datasets show that DFGR outperforms DLRM, which serves as the industrial online baseline with extensive feature engineering, as well as Meta's HSTU and other common recommendation models such as DIN, DCN, DIEN, and DeepFM. Furthermore, we investigate optimal parameter allocation strategies under computational constraints, establishing DFGR as an efficient and effective next-generation generative ranking paradigm.

Figures

Figures reproduced from arXiv: 2505.16752 by the authors.

Figure 1
Figure 1. MetaGR’s Generative Ranking Architecture [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Model Architecture of Single-Flow Generative Ranking Network. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Training Sample Construction Method under the [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Model Architecture of Dual-Flow Generative Ranking Network. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Session-Aware Cross Triangle Mask. mask matrix as shown in [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: DFGR vs DFGR-MOA Performance. truncated to a maximum length of 4K, (2) identical model archi￾tecture and training hyperparameters were employed, and (3) the DNN output layers were partitioned into channels in the full do￾main model to preserve the consistency of the va…
Figure 7
Figure 7. Figure 7: Constant vs Decay Performance. 0.0 100.0 200.0 300.0 400.0 500.0 600.0 700.0 800.0 Computational Complexity (PetaFLOPs) 0.705 0.710 0.715 0.720 0.725 0.730 G-AUC Layer 1, Dim 360 Layer 2, Dim 216 Layer 4, Dim 120 Layer 8, Dim 64 Layer 16, Dim 32 Max G-AUC [PITH_FULL_I…
Figure 8
Figure 8. Figure 8: Performance Comparison of Different Model Pa [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]

Discussion (0). Sign in 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. GR-LLMs: Recent Advances in Generative Recommendation Based on Large Language Models

    cs.IR 2025-07 unverdicted novelty 3.0 of 10

    A survey of LLM-based generative recommendation systems, covering application settings, training pipelines, industrial deployment challenges, and future directions.

Reference graph

Works this paper leans on

38 extracted references · 4 canonical work pages · cited by 1 Pith paper

  1. [1]

    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

  2. [2]

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

  3. [3]

    Stefan Elfwing, Eiji Uchibe, and Kenji Doya. 2018. Sigmoid-weighted linear units for neural network function approximation in reinforcement learning. Neural networks 107 (2018), 3–11

    2018

  4. [4]

    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)

    arXiv 2019

  5. [5]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783 (2024)

    arXiv 2024

  6. [6]

    Huifeng Guo, Ruiming Tang, Yunming Ye, Zhenguo Li, and Xiuqiang He. 2017. DeepFM: a factorization-machine based neural network for CTR prediction.arXiv preprint arXiv:1703.04247 (2017)

    arXiv 2017

  7. [7]

    Xiangnan He and Tat-Seng Chua. 2017. Neural factorization machines for sparse predictive analytics. In Proceedings of the 40th International ACM SIGIR conference on Research and Development in Information Retrieval . 355–364

    2017

  8. [8]

    Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, et al. 2022. Training compute-optimal large language models. arXiv preprint arXiv:2203.15556 (2022)

    arXiv 2022

Show all 38 references
  1. [9]

    Shengding Hu, Yuge Tu, Xu Han, Ganqu Cui, Chaoqun He, Weilin Zhao, Xiang Long, Zhi Zheng, Yewei Fang, Yuxiang Huang, et al. 2024. MiniCPM: Unveiling the Potential of Small Language Models with Scalable Training Strategies. In First Conference on Language Modeling

    2024

  2. [10]

    Weizhe Hua, Zihang Dai, Hanxiao Liu, and Quoc Le. 2022. Transformer quality in linear time. In International conference on machine learning. PMLR, 9099–9117

    2022

  3. [11]

    Tongwen Huang, Zhiqi Zhang, and Junlin Zhang. 2019. FiBiNET: combining fea- ture importance and bilinear feature interaction for click-through rate prediction. In Proceedings of the 13th ACM conference on recommender systems . 169–177

    2019

  4. [12]

    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

  5. [13]

    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

  6. [14]

    Bin Liu, Ruiming Tang, Yingzhi Chen, Jinkai Yu, Huifeng Guo, and Yuzhou Zhang

  7. [15]

    Jiaqi Ma, Zhe Zhao, Xinyang Yi, Jilin Chen, Lichan Hong, and Ed H Chi. 2018. Modeling task relationships in multi-task learning with multi-gate mixture-of- experts. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining . 1930–1939

    2018

  8. [16]

    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

  9. [17]

    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

  10. [18]

    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

  11. [19]

    Changhua Pei Shanshan Lv Tao Zhuang Qiwei Chen, Yue Xu and Junfeng Ge. 2022. Efficient long sequential user data modeling for click-through rate prediction. arXiv preprint arXiv:2209.12212 (2022)

    2022 arXiv

  12. [20]

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

  13. [21]

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research 21, 140 (2020), 1–67

    2020

  14. [22]

    Ying Shan, T Ryan Hoens, Jian Jiao, Haijing Wang, Dong Yu, and JC Mao. 2016. Deep crossing: Web-scale modeling without manually crafted combinatorial features. In Proceedings of the 22nd ACM SIGKDD international conference on Action is All You Need: Dual-Flow Generative Rankin...

    2016

  15. [23]

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

  16. [24]

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yas- mine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhos- ale, et al. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023)

    2023 arXiv

  17. [25]

    Ruoxi Wang, Rakesh Shivanna, Derek Cheng, Sagar Jain, Dong Lin, Lichan Hong, and Ed Chi. 2021. Dcn v2: Improved deep & cross network and practical lessons for web-scale learning to rank systems. In Proceedings of the web conference 2021 . 1785–1797

    2021

  18. [26]

    Ying Wen, Tianyao Chen, Jun Wang, and Weinan Zhang. 2019. Pairwise multi- layer nets for learning distributed representation of multi-field categorical data. In Proceedings of the 1st international workshop on deep learning practice for high-dimensional sparse data. 1–8

    2019

  19. [27]

    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

  20. [28]

    Jun Xiao, Hao Ye, Xiangnan He, Hanwang Zhang, Fei Wu, and Tat-Seng Chua

  21. [29]

    Jiaqi Zhai, Lucy Liao, Xing Liu, Yueming Wang, Rui Li, Xuan Cao, Leon Gao, Zhao- jie Gong, Fangda Gu, Michael He, et al. 2024. Actions speak louder than words: Trillion-parameter sequential transducers for generative recommendations.arXiv preprint arXiv:2402.17152 (2024)

    2024 arXiv

  22. [30]

    Buyun Zhang, Liang Luo, Xi Liu, Jay Li, Zeliang Chen, Weilin Zhang, Xiaohan Wei, Yuchen Hao, Michael Tsang, Wenjun Wang, et al. 2022. DHEN: A deep and hierarchical ensemble network for large-scale click-through rate prediction. arXiv preprint arXiv:2203.11014 (2022)

    2022 arXiv

  23. [31]

    Xue Xia, Pong Eksombatchai, Nikil Pancha, Dhruvil Deven Badani, Po-Wei Wang, Neng Gu, Saurabh Vishwas Joshi, Nazanin Farahpour, Zhiyuan Zhang, and An- drew Zhai. 2023. Transact: Transformer-based realtime user action model for recommendation at pinterest. In Proceedings of the...

    2023

  24. [32]

    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

  25. [33]

    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

  26. [36]

    Xiaofeng Zhang, Yikang Shen, Zeyu Huang, Jie Zhou, Wenge Rong, and Zhang Xiong. 2022. Mixture of attention heads: Selecting attention heads per token. arXiv preprint arXiv:2210.05144 (2022)

    2022 arXiv

  27. [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

  28. [2017]

    arXiv preprint arXiv:1708.04617 (2017)

    Attentional factorization machines: Learning the weight of feature interac- tions via attention networks. arXiv preprint arXiv:1708.04617 (2017)

    2017 arXiv

  29. [2019]

    In The World Wide Web Conference

    Feature generation by convolutional neural network for click-through rate prediction. In The World Wide Web Conference. 1119–1129

  30. [2021]

    arXiv preprint arXiv:2108.04468 (2021)

    End-to-End User Behavior Retrieval in Click-Through Rate Prediction Model. arXiv preprint arXiv:2108.04468 (2021)

    2021 arXiv

Pith tools

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