Pith. sign in

REVIEW 4 major objections 5 minor 3 cited by

EMER, an end-to-end ensemble ranking framework that treats each recommendation request as a single transformer-readable candidate set with self-evolving multi-objective losses, beats the deployed fusion formula and the UREM baseline in live

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

EMER is an end-to-end ensemble ranking framework for short video recommendation; at Kuaishou it lifted app stay time by 1.39% while improving most interaction metrics.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A credible industrial report on an end-to-end ensemble ranking framework with real deployment gains, but the evidence is thinner than the claims — needs statistical rigor and a real derivation for IPUT. the 4 major comments →

arxiv 2508.05093 v2 pith:ZZLOUVQL submitted 2025-08-07 cs.IR

An End-to-End Multi-objective Ensemble Ranking Framework for Video Recommendation

Tiantian He , Minzhi Xie , Runtong Li , Xiaoxiao Xu , Jiaqi Yu , Zixiu Wang , Lantao Hu , Han Li
show 1 more author
Kun Gai
This is my paper
classification cs.IR
keywords recommender systemsensemble rankingmulti-objective optimizationtransformerAUC surrogate lossself-evolving weightsIPUToffline-online consistency
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

The paper argues that the final ranking stage of a short-video recommender—the step that fuses many predicted user-satisfaction scores into one order—should itself be learned end-to-end, instead of being set by a hand-tuned formula. To do this it introduces EMER, which groups all candidates in a request into one sample, encodes their relative positions, and uses a transformer to compare them. Supervision comes from two complementary sources: a pairwise 'relative advantage' loss on real post-exposure feedback, and multi-objective AUC-style losses on dense model predictions (watch time, click, like, share, etc.) whose weights are adjusted by an 'Advantage Evaluator' that compares the current model with its own previous version. A new offline target, IPUT (interaction probability per unit time), is designed to remove watch-time confounding so that offline AUC gains predict online behavior. In live A/B tests on a large short-video platform, EMER improves overall app stay time by about 1.39% and 7-day lifetime by about 0.20% over the deployed fusion formula.

Core claim

The central claim is that multi-objective ensemble ranking can be cast as a supervised learning problem without a single ground-truth label, and that the resulting end-to-end model outperforms both a manually designed fusion formula and an unsupervised ensemble baseline. EMER packages each recommendation request as one training instance containing all candidate videos, exposed or not; a transformer reads those candidates plus normalized-rank features to produce comparative scores. The loss combines a posterior pairwise ranking term built on the observed hierarchy of user feedback with multiple pairwise AUC surrogate losses over prior model predictions, each weighted by a dynamic coefficient

What carries the argument

The load-bearing mechanism is the request-level comparison model: all candidates of one recommendation request—exposed and unexposed—form a single training sample; per-item NormalizedRanks (rank divided by candidate count) are fed alongside features into a transformer that scores the entire set. Around that core sit three supporting pieces: the 'Relative Advantage Satisfaction' pairwise loss over posterior feedback; the multi-objective AUC surrogate losses over prior predictions $P_{xtrs}$, with per-objective weights produced by an Advantage Evaluator that computes, per request, the ratio of a ranking metric ($DCG@K$) of the current model to the serving model; and the IPUT reformulation $p_{

Load-bearing premise

The assumption is that the multi-task model's predicted satisfaction signals (watch time, click, like, share, etc.), used as dense supervision for every candidate including ones never shown, carry no exposure or popularity bias of their own—so that EMER learns genuine user satisfaction rather than imitating the biases of the teacher model that produced those predictions.

What would settle it

Look at how EMER actually ranks unexposed items versus exposed ones in a logged dataset where the true outcome for the unexposed item can be inferred from a randomized interleaving or a policy evaluation. If, after swapping the top-ranked unexposed item with the top-ranked exposed item, user engagement follows the original exposure order rather than EMER's preference order, the claim that the prior signals are unbiased supervision would be falsified. A cheaper indicator: train EMER twice, once with real $P_{xtrs}$ and once with $P_{xtrs}$ that have been explicitly re-ranked to match exposure o

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

If this is right

  • Any ranking stage that currently uses a scalarized hand-tuned formula can be replaced by a learned comparator, trained with no manually assigned loss weights.
  • Because supervision includes dense model predictions for unexposed items, EMER can rank candidates it has never seen feedback on, in principle mitigating exposure bias—if the predictions themselves are unbiased.
  • The self-evolving weight scheme removes manual hyperparameter tuning for multi-objective tradeoffs, and the DCG@K-based Advantage Evaluator offers a concrete, plug-in metric for dynamic weighting.
  • The IPUT target creates a template for other ratio-based offline objectives: divide any cumulative engagement count by predicted exposure time to make offline optimization consistent with online session-level goals.
  • The observed gains—1.39% stay time and roughly 0.20% LT7 at the scale of hundreds of millions of daily active users—imply that even small offline GAUC improvements in ensemble ranking translate into meaningful product-level engagement.

Where Pith is reading between the lines

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

  • If the underlying multi-task model that produces $P_{xtr}$ is itself exposure-biased, then the 'unexposed' supervision may mostly distill the teacher's biases; a strong test would be to train EMER on deliberately corrupted $P_{xtrs}$ and see whether offline GAUC gains persist.
  • The per-request transformer formulation resembles a set-to-sequence scoring model; one could extend it to jointly model the order of the whole list (listwise objectives) rather than pairwise surrogates, which might capture position interactions beyond relative advantage.
  • The IPUT reformulation suggests a broader principle: any offline metric that counts events in a session can be made density-based by dividing by time, and EMER's online results suggest this is a generally useful correction for session-level engagement objectives.
  • The self-evolving weight mechanism is a form of online meta-learning (comparing the updated model against the serving model per request); the same idea could be used to adapt other multi-task recommendation losses without storing checkpoints.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes EMER, an end-to-end multi-objective ensemble ranking framework for short-video recommendation. EMER replaces a manually tuned fusion formula with a transformer-based model that takes all candidates in a request as input, adds normalized rank features, and is trained with a pairwise posterior loss (derived from 'Relative Advantage Satisfaction') plus multi-objective prior AUC losses over pxtr signals (pvtr, pctr, pltr, etc.). A self-evolving scheme adjusts per-objective loss weights using an Advantage Evaluator that compares current and previous model performance. The paper also introduces IPUT (Interaction Probability per Unit Time) as an offline evaluation/training target intended to resolve the 'decoupling paradox' between offline interaction AUC and online interaction rate. The authors report offline GAUC results and online A/B tests in two Kuaishou scenarios, claiming substantial gains in LT7, App Stay Time, watch time, and several interaction metrics (e.g., +1.392% AppStayTime in Scenario#1).

Significance. If the empirical claims are valid, the paper makes a practical contribution: it demonstrates at industrial scale that a learned, candidate-comparison-aware ensemble can replace heuristic fusion and improve multiple user-satisfaction metrics simultaneously. The deployment in a major short-video platform and the use of very large-scale training data are strengths. The paper also identifies a real problem (offline-online inconsistency in multi-objective ranking) and proposes a concrete, if simple, correction (IPUT). However, the current evidence base is not sufficient to establish the central claims: the online A/B results lack statistical characterization, and the offline GAUC analyses are partly circular because the same pxtr signals serve as both training labels and evaluation targets. The causal justification for IPUT is asserted rather than derived. These gaps mean the significance currently rests on an industrial claim that is not fully verifiable from the manuscript.

major comments (4)
  1. [Experiments, Table 1] The central claim that 'EMER significantly outperforms baseline approaches' rests on Table 1, but no confidence intervals, p-values, experiment duration, user allocation, or metric pre-registration are reported. There are 16 comparisons (8 metrics × 2 scenarios); several metrics move negatively (e.g., Like -0.270% in Scenario#2). Under multiple testing, the reported lifts may be within noise. Without a statistical framework, the A/B results cannot be distinguished from chance, so the paper's main empirical conclusion is unestablished.
  2. [Learning Objectives; Tables 2–6] The offline GAUC evaluation is largely circular. Table 2 measures agreement between EMER's output ranking and the pxtr rankings that are directly used as pair labels in Eqs. (2)–(5). High consistency with pxtr is expected when the model is trained to reproduce those orderings. Similarly, the IPUT variants in Table 5 use pxtr_IPUT as a label that EMER is trained against. The ablation studies (Tables 3–4) compare variants trained with different label sets, so differences are also expected from the objective change. None of these analyses provides an independent offline signal of user satisfaction. This does not disprove the method, but it means the offline evidence cannot substitute for rigorous A/B statistics.
  3. [Evaluation methodology, Eq. (10)] The IPUT definition pxtr_IPUT = pxtr / p_watchtime is presented as a causal adjustment that 'eliminates the confounding factor of watch time inherent in interactions' using the backdoor criterion. However, no causal diagram, no definition of the variable set, and no derivation of the backdoor adjustment are provided. The numerical example (three candidates) is an illustration, not a proof. As written, Eq. (10) is a ratio of two model predictions, and its causal semantics are unclear. If the intent is that p_watchtime is a confounder, the backdoor criterion would require conditioning, not simple division, and the paper gives no argument that this ratio identifies the desired interaction intensity. This is load-bearing because IPUT is a core contribution and underpins the claimed offline-online consistency.
  4. [Self-Evolving Optimization Scheme, Eqs. (8)–(9)] The self-evolving weight formula w_pxtr = Metric(f(Θ_t−1))/Metric(f(Θ_t)) is introduced as a dynamic, per-request, personalized weighting scheme. Several issues are load-bearing. First, Eq. (9) is written with a sum/dataset index (x_i, y_i), indicating a dataset-level scalar, not a per-request weight; the text claims per-request personalization but the equation does not implement it. Second, if Metric is a higher-is-better ranking measure (e.g., DCG@K), an improving objective receives a smaller weight (ratio < 1), which is counterintuitive for a scheme meant to 'constrain the model towards comprehensively superior AUC performance'; the paper does not explain why this direction is correct or how it resolves objective conflicts. Third, no convergence, stability, or non-degradation argument is provided. The design may be defensible, but as presented it is an ad hoc heuristic, not a justifie
minor comments (5)
  1. [Appendix, Eq. (13)] DCG@K appears in Eq. (13) with an unspecified K. The choice of K is a free parameter; the paper should state the value used in Table 6 and justify the sensitivity.
  2. [Experimental Setup, training scale] The paper says 'over ten billion samples per day, with each sample containing approximately 500 candidate videos.' Clarify whether each 'sample' is a request-level aggregated sample or a single item impression; the architecture in Figure 2 has a candidate size dimension, so the ambiguity matters.
  3. [Table 4 / Ablation text] The ablation paragraph gives numbers for EMER-NoEvolve ('VideoView -2.347%, Forward -8.418%, Comment -8.109%') that do not match Table 4 (VideoView -1.303%, Forward -8.088%, Comment -7.414%). Please reconcile.
  4. [Figure 4 caption] The caption of Figure 4 mentions 'EMER without MSE,' but no MSE loss is defined in the method. This appears to be a typo and should be corrected to refer to the self-evolving scheme or the static-weight variant.
  5. [Related Work, references] The related-work section mentions IntEL and UREM, but not a formal comparison or discussion of why UREM is a reasonable supervised-ensemble baseline beyond being 'unsupervised.' A brief justification would help.

Circularity Check

2 steps flagged

Offline GAUC/IPUT evaluations reuse training labels, so they are self-consistency checks; online A/B is non-circular but statistically under-reported.

specific steps
  1. fitted input called prediction [Experiments, Overall Performance (Table 2); Learning Objectives, Eqs. (2)-(5)]
    "To systematically evaluate these ranking methods, we conduct an analysis of rank consistency between the rankings generated by ensemble methods and those derived from pxtrs. ... Table 2 indicates that our method has better consistency with most of the input pxtrs."

    The pxtr rankings used as evaluation targets in Table 2 are exactly the supervision labels in Eqs. (2)-(5): L_prior = -1/N * sum_{pxtr} w_pxtr * L_pxtr, where L_pxtr is the pairwise loss over y^{pxtr}_{ij} = 1 if pxtr_i > pxtr_j. Thus EMER is trained to maximize agreement with these pxtr rankings, and reporting GAUC against those same pxtr rankings is a self-consistency check, not an independent test. It mainly confirms that the model fit its own training targets.

  2. fitted input called prediction [Evaluation methodology, Eq. (10); Experiments, Ablation of Interaction Probability per Unit Time, Table 5]
    "As shown in Table 5, EMER-NoIPUT suffers from a classic decoupling paradox: significant offline improvements in the AUC of pltr and pwtr fail to translate into positive online A/B test results. In contrast, positive offline gains of pxtrIPUT directly align with positive online results."

    IPUT is a reformulated prior signal (Eq. 10: pxtr_IPUT = pxtr / p_watchtime), and as with other Pxtrs, EMER's L_prior (Eq. 5) is trained to maximize the pairwise ranking of pxtr_IPUT. Therefore EMER's higher offline GAUC on pltrIPUT/pwtrIPUT/pcmtrIPUT/pftrIPUT relative to EMER-NoIPUT is by construction: the IPUT variant is directly optimized for those labels, while the NoIPUT variant is not. This offline comparison is a fitted-input check, not independent evidence of offline-online consistency.

full rationale

The paper's central claim—EMER beats the deployed fusion formula and UREM online—rests on Table 1 A/B lifts, which are genuine external comparisons and not circular. I found no load-bearing self-citation chain, no imported uniqueness theorem, and no ansatz smuggled in via citation. However, the offline evidence is partially circular. Table 2's GAUC measures agreement with pxtr rankings that Eqs. (2)-(5) use as training labels, so it mostly confirms that EMER fit its own targets. Table 5's offline IPUT GAUC similarly compares a model trained on IPUT with one not trained on IPUT, which is a fitted-input comparison. The 'prior signals are unbiased' assumption is an unverified modeling claim, not a circular step. The online A/B results are non-circular but are reported without confidence intervals, p-values, or multiple-comparison details; that is a statistical robustness concern, not a circularity one. Overall, the derivation is not equivalent to its inputs, but the offline evaluation sections should be read as consistency checks rather than independent verification, yielding a score of 6 for partial circularity.

Axiom & Free-Parameter Ledger

3 free parameters · 6 axioms · 3 invented entities

The framework rests on several unverified modeling choices: prior model predictions treated as unbiased labels, a causal intervention claimed but not formalized, a self-evolving weight formula with no theoretical grounding, and an ordinal posterior label definition. These are the prices paid for an industrial deployment paper that does not release data or code.

free parameters (3)
  • K in DCG@K (Advantage Evaluator metric) = not reported
    The top-K cutoff is chosen in the Metric Study (Table 6) based on offline GAUC; it controls the dynamic loss weights (Eq 8, 9, 13) and therefore the final ranking.
  • Posterior/prior loss combination coefficient = 1.0 (implicit)
    Equation 6 sums L_posterior and L_prior with no learned or tuned weight; the implicit 1:1 weighting is a modeling choice not justified empirically.
  • Dynamic objective weights w_pxtr = computed per request from metric ratios
    Although not manually fitted, the AE formula (Eq 9) is an ad hoc construction; its scale and behavior are not grounded in any multi-objective trade-off principle.
axioms (6)
  • domain assumption Each prior signal pxtr reflects a unique facet of user satisfaction, and improving each individually improves overall satisfaction.
    Stated in 'Enhancing multi-dimension ranking capability' and underpins the multi-objective loss (Eq 5).
  • ad hoc to paper The backdoor criterion applies and normalization pxtr/p_watchtime removes confounding of watch time from interactions.
    Invoked in 'Evaluation methodology' Eq 10, no causal diagram or derivation is given.
  • ad hoc to paper The ratio of previous/current model metrics used as dynamic weight improves multi-objective balance.
    Equation 8-9 defines w_pxtr = Metric(theta_t-1)/Metric(theta_t), but no proof or mechanism is provided that this weighting converges to a desirable trade-off.
  • domain assumption Unexposed candidate items are usable as comparison items with prior-model predictions as labels, eliminating exposure bias.
    Sample Organization claims grouping exposed and unexposed items overcomes exposure bias; this assumes prior predictions are unbiased surrogates.
  • domain assumption Online A/B metrics are attributable to the ranking change with no interference or novelty effects.
    Table 1 reports relative improvements without confidence intervals; standard A/B assumptions are unstated.
  • domain assumption Pairwise ordering 'Many Positives > Single Positive > No Positive' approximates user satisfaction.
    Modeling Relative Advantage Satisfaction Labels; no calibration or validation of this ordinal assumption.
invented entities (3)
  • Relative Advantage Satisfaction label no independent evidence
    purpose: Provides pairwise supervision for ensemble ranking based on posterior feedback categories.
    The label ordering (Many > Single > No Positive) is invented and validated only through internal ablations; no external falsifiable handle.
  • IPUT (Interaction Probability per Unit Time) no independent evidence
    purpose: Offline-online consistent target defined as pxtr/p_watchtime.
    It is a redefinition of existing predictions with no independent measurement or external validation of the causal claim.
  • Advantage Evaluator no independent evidence
    purpose: Computes dynamic loss weights from previous/current model metric ratios.
    A novel component whose behavior is demonstrated only on this system; no independent evidence outside the paper.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of An End-to-End Multi-objective Ensemble Ranking Framework for Video Recommendation." pith.science (2026). https://pith.science/paper/ZZLOUVQL

@misc{pith2026250805093,
  author       = {Pith},
  title        = {Pith review of: An End-to-End Multi-objective Ensemble Ranking Framework for Video Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZZLOUVQL}},
  note         = {Machine review of arXiv:2508.05093}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

We propose a novel End-to-end Multi-objective Ensemble Ranking framework (EMER) for the multi-objective ensemble ranking module, which is the most critical component of the short video recommendation system. EMER enhances personalization by replacing manually-designed heuristic formulas with an end-to-end modeling paradigm. EMER introduces a meticulously designed loss function to address the fundamental challenge of defining effective supervision for ensemble ranking, where no single ground-truth signal can fully capture user satisfaction. Moreover, EMER introduces novel sample organization method and transformer-based network architecture to capture the comparative relationships among candidates, which are critical for effective ranking. Additionally, we have proposed an offline-online consistent evaluation system to enhance the efficiency of offline model optimization, which is an established yet persistent challenge within the multi-objective ranking domain in industry. Abundant empirical tests are conducted on a real industrial dataset, and the results well demonstrate the effectiveness of our proposed framework. In addition, our framework has been deployed in the primary scenarios of Kuaishou, a short video recommendation platform with hundreds of millions of daily active users, achieving a 1.39% increase in overall App Stay Time and a 0.196% increase in 7-day user Lifetime(LT7), which are substantial improvements.

Figures

Figures reproduced from arXiv: 2508.05093 by Han Li, Jiaqi Yu, Kun Gai, Lantao Hu, Minzhi Xie, Runtong Li, Tiantian He, Xiaoxiao Xu, Zixiu Wang.

Figure 1
Figure 1. Figure 1: Heatmaps of normalized user behavior indicators. (left) [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: An illustration of the EMER framework. Online Serving Model Advantage Evaluator Training Model Labeled Data Metric Evaluation Metric Evaluation Dynamic Loss Weight [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: End-to-end pipeline with online self-evolution. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Multi-Objective Loss Distribution Comparison. EMER [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Relative Improvement in LT7 metrics over Fusion [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 3 Pith papers

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

  1. Uncertainty as Remedy: Mitigating Satisfaction Label Bias in Short Video Multi-Objective Ensemble Ranking

    cs.IR 2026-07 conditional novelty 6.0

    UAME models each predicted satisfaction score as a Gaussian with learned variance, then uses that variance to up-weight conflicting item pairs in a probabilistic pairwise ranking loss, improving multi-objective short-...

  2. UMRE: A Unified Monotonic Transformation for Ranking Ensemble in Recommender Systems

    cs.IR 2025-08 unverdicted novelty 5.0

    UMRE learns strictly monotonic score transformations and user-personalized fusion weights for ensemble ranking, coordinated by an adaptive Pareto task-weighting strategy.

  3. SalesLoop: Reinforcement Learning from Performance Feedback for Sales Lead Ranking

    cs.LG 2026-07 conditional novelty 4.0

    A monthly closed-loop RL system for ranking sales leads, using position- and speed-weighted conversion rewards and a listwise GRPO-style loss, reports +7.9% NDCG@K offline and +4.7%/+8.7% conversion lift in a 160-day ...

Reference graph

Works this paper leans on

27 extracted references · 17 canonical work pages · cited by 3 Pith papers · 1 internal anchor

  1. [1]

    Michał Bałchanowski and Urszula Boryczka. 2022. Aggregation of rankings using metaheuristics in recommendation systems. Electronics 11, 3 (2022), 369

    work page 2022

  2. [2]

    Michał Bałchanowski and Urszula Boryczka. 2022. Collaborative rank aggregation in recommendation systems. Procedia computer science 207 (2022), 2213–2222

    work page 2022

  3. [3]

    Toon Calders and Szymon Jaroszewicz. 2007. Efficient AUC optimization for classification. In European conference on principles of data mining and knowledge discovery. Springer, 42–53

    work page 2007

  4. [4]

    Yukuo Cen, Jianwei Zhang, Xu Zou, Chang Zhou, Hongxia Yang, and Jie Tang

  5. [5]

    Konstantina Christakopoulou, Can Xu, Sai Zhang, Sriraj Badam, Trevor Pot- ter, Daniel Li, Hao Wan, Xinyang Yi, Ya Le, Chris Berg, et al . 2022. Reward Shaping for User Satisfaction in a REINFORCE Recommender. arXiv preprint arXiv:2209.15166 (2022)

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  6. [6]

    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)

    Pith/arXiv arXiv 2014

  7. [7]

    Corinna Cortes and Mehryar Mohri. 2003. AUC optimization vs. error rate minimization. Advances in neural information processing systems 16 (2003)

    work page 2003

  8. [8]

    Ronald Fagin, Ravi Kumar, and Dandapani Sivakumar. 2003. Efficient similarity search and classification via rank aggregation. In Proceedings of the 2003 ACM SIGMOD international conference on Management of data. 301–312

    work page 2003

  9. [9]

    Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk

  10. [10]

    Chao Li, Zhiyuan Liu, Mengmeng Wu, Yuchi Xu, Huan Zhao, Pipei Huang, Guoliang Kang, Qiwei Chen, Wei Li, and Dik Lun Lee. 2019. Multi-interest network with dynamic routing for recommendation at Tmall. InProceedings of the 28th ACM international conference on information and knowledge management. 2615–2623

    2019

  11. [11]

    Jiayu Li, Peijie Sun, Zhefan Wang, Weizhi Ma, Yangkun Li, Min Zhang, Zhoutian Feng, and Daiyue Xue. 2023. Intent-aware Ranking Ensemble for Personalized Recommendation (SIGIR ’23). Association for Computing Machinery, New York, NY , USA, 1004–1013. https://doi.org/10.1145/3539618.3591702

    arXiv 2023

  12. [12]

    Fuyu Lv, Taiwei Jin, Changlong Yu, Fei Sun, Quan Lin, Keping Yang, and Wil- fred Ng. 2019. SDM: Sequential deep matching model for online large-scale recommender system. In Proceedings of the 28th ACM international conference on information and knowledge management. 2635–2643

    work page 2019

  13. [13]

    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

  14. [14]

    Samuel Oliveira, Victor Diniz, Anisio Lacerda, and Gisele L Pappa. 2016. Evolu- tionary rank aggregation for recommender systems. In 2016 IEEE Congress on Evolutionary Computation (CEC). IEEE, 255–262

    work page 2016

  15. [15]

    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

    work page 2019

  16. [16]

    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 Information & Knowledge Management. 2685–2692

    work page 2020

  17. [17]

    Qiaoyu Tan, Jianwei Zhang, Ninghao Liu, Xiao Huang, Hongxia Yang, Jingren Zhou, and Xia Hu. 2021. Dynamic memory based attention network for sequential recommendation. In Proceedings of the AAAI conference on artificial intelligence, V ol. 35. 4384–4392

    work page 2021

  18. [18]

    Hongyan Tang, Junning Liu, Ming Zhao, and Xudong Gong. 2020. Progressive layered extraction (ple): A novel multi-task learning (mtl) model for personalized recommendations. In Proceedings of the 14th ACM conference on recommender systems. 269–278

    2020

  19. [19]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention Is All You Need. arXiv:1706.03762 [cs.CL]

    Pith/arXiv arXiv 2017

  20. [20]

    Jing Yan, Liu Jiang, Jianfei Cui, Zhichen Zhao, Xingyan Bin, Feng Zhang, and Zuotao Liu. 2024. Trinity: Syncretizing Multi-/Long-Tail/Long-Term Interests All in One. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 6095–6104

    work page 2024

  21. [21]

    Congrui Yi, David Zumwalt, Zijian Ni, and Shreya Chakrabarti. 2023. Progressive horizon learning: Adaptive long term optimization for personalized recommen- dation. In Proceedings of the 17th ACM Conference on Recommender Systems. 940–946

    work page 2023

  22. [22]

    Wenhui Yu, Bingqi Liu, Bin Xia, Xiaoxiao Xu, Ying Chen, Yongchang Li, and Lantao Hu. 2024. Unsupervised Ranking Ensemble Model for Recommendation. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD 2024, Barcelona, Spain, August 25-29, 2024 , Ricardo Baeza-Yates and Francesco Bonchi (Eds.). ACM, 6181–6189. https:/...

    arXiv 2024

  23. [23]

    Qihua Zhang, Junning Liu, Yuzhuo Dai, Yiyan Qi, Yifan Yuan, Kunlun Zheng, Fan Huang, and Xianfeng Tan. 2022. Multi-task fusion via reinforcement learning for long-term user satisfaction in recommender systems. In Proceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining. 4510–4520

    work page 2022

  24. [24]

    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, V ol. 33. 5941–5948

    work page 2019

  25. [25]

    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 inter- national conference on knowledge discovery & data mining. 1059–1068. APPENDIX • 𝐻𝑖𝑡𝑅𝑎𝑡𝑒 @𝐾: The intersection of the top-K items from th...

    work page 2018

  26. [2015]

    arXiv preprint arXiv:1511.06939 (2015)

    Session-based recommendations with recurrent neural networks. arXiv preprint arXiv:1511.06939 (2015)

    Pith/arXiv arXiv 2015

  27. [2020]

    Controllable multi-interest framework for recommendation. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & Conference’17, July 2017, Washington, DC, USA Tiantian He ∗, Minzhi Xie∗, Runtong Li∗, Xiaoxiao Xu∗, Jiaqi Y u∗, Zixiu Wang∗, Lantao Hu∗, Han Li∗, and Kun Gai∗ data mining. 2942–2951

    work page 2017

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.