Pith. sign in

REVIEW 3 major objections 5 minor 38 references

The paper claims that injecting explicit content-type context into gating and expert representations lets one unified ranker serve a heterogeneous feed, improving positive engagement while reversing negative-engagement regression and closin

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

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

T0 review · deepseek-v4-flash

2026-08-01 05:21 UTC pith:SYTTZ6WU

load-bearing objection A solid industrial case study with a plausible deployment story, but the claim that heterogeneity context drives the gains is not actually supported by the ablations; still worth refereeing. the 3 major comments →

arxiv 2607.27577 v1 pith:SYTTZ6WU submitted 2026-07-30 cs.IR cs.LG

Heterogeneous Ranking in Industrial-Scale Recommender Systems: A Case Study

Di Bai , Jintao Liu , Zhenwei Tang , Peifan Wu , Nada Al-Thawr , Luoshu Wang This is my paper
classification cs.IR cs.LG
keywords recommender systemsmulti-task learningmixture-of-expertsheterogeneous rankingnegative transfercross-type evaluationproduction deploymentGoogle Discover
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper is a production case study arguing that a single multi-task ranker can serve a feed mixing web articles, videos, and UGC posts without suffering negative transfer or majority bias. The proposed HA-MoE architecture injects an explicit heterogeneity vector — content type, followed-creator status, shoppability, AI-summary — into both the gating network and a late-layer affine modulation of each expert. Offline, this raises DL-AUC for positive and negative engagement tasks and narrows the cross-type xAUC gap between articles and videos from 0.141 to 0.060. Online A/B tests on live traffic show lifts in daily active users, viewed impressions, scroll depth, and diverse engagement. A sympathetic reader would take away that explicit heterogeneity conditioning, rather than more parameters or separate per-format models, is the key to unified heterogeneous ranking.

Core claim

On the paper's own terms, the central discovery is that the negative-transfer and minority-bias problems of a shared-backbone multi-task ranker can be broken by making the model explicitly aware of what kind of content it is ranking. HA-MoE feeds the encoded heterogeneity signals h into a multi-gate mixture-of-experts at two points: the gating network's fusion input (a simple concatenation [x∥h] in the initial deployment) and a Heterogeneity-Driven Linear Modulation (HDLM) layer that scales and shifts each expert's final representation. The result is that the model can route different content types to different experts and adjust representations contextually — protecting videos from being bu

What carries the argument

The load-bearing object is the explicit heterogeneity vector h — categorical and contextual signals such as content_type, followed-creator, shoppable, and AI-summary — used in two places. First, HA-Gating: each task's softmax gate over N=4 shared experts takes a fusion Φ(x,h)=[x∥h] as input, so expert selection can depend on content type. Second, HDLM: each expert's last layer is modulated by a learned affine transform γ_n(h)⊙E_n(x)+β_n(h), letting the same expert behave differently across content types. The evaluation machinery is DL-AUC = λ·Micro-AUC + (1−λ)·Macro-xAUC, which combines global ranking quality with an unweighted average of cross-type xAUC pairs to catch majority score inflati

Load-bearing premise

The improvement is attributed to the explicit heterogeneity signals h; if those signals are already recoverable from the dense features, or if h is an incomplete proxy for true content structure, the gains could come from added model capacity or incidental conditioning rather than from heterogeneity adaptation.

What would settle it

Retrain HA-MoE with the heterogeneity vector h replaced by a constant (or by random noise) while keeping the same parameter count and training setup; if DL-AUC and the article/video cross-type gap remain close to the reported values, the paper's central attribution to explicit heterogeneity signals is unsupported. A second check: probe whether a model trained with dense features only can predict h, which would indicate the signals are redundant.

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

If this is right

  • A unified heterogeneous ranker can replace the common practice of separate per-format rankers merged late, since one HA-MoE model improved both macro task groups at once.
  • Explicitly conditioning on content type protects minority segments: the Vid+ vs. Web− xAUC rose from 0.830 to 0.900, narrowing the cross-type gap from 0.141 to 0.060.
  • The gains translate online: live A/B tests showed +0.22% DAU, +0.48% viewed impressions, +0.34% scroll depth, and +0.36–0.54% diversity metrics with no regressions.
  • Production constraints need not block specialization: HA-MoE adds under 5% model size, comparable training speed, and under 0.5% serving latency.
  • The DL-AUC metric provides a single automated decision signal that penalizes dominant-score inflation, making it usable as a deployment guardrail.

Where Pith is reading between the lines

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

  • Inference: If h is the active ingredient, the same cheap modulation pattern could be applied to other axes of variation — user context, device, time, or business objective — suggesting a general recipe for unifying disparate segments in one model.
  • Inference: The macro-xAUC component of DL-AUC is a provider-fairness lens; the paper does not develop it, but the same formula could be reported for creator categories or publishers, not just content types.
  • Inference: LENS/PIEM could be used as a general collapse detector for any MoE recommender, because it is label-free and computes on gating profiles; the paper demonstrates it on one feed but the mechanism is architecture-generic.
  • Inference: A testable extension is to replace the concatenation fusion Φ with a learned attention over h, or to add h at every expert layer; the paper leaves this as future work and currently applies HDLM only to the final expert layer.

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

3 major / 5 minor

Summary. This paper presents a production case study of HA-MoE, a mixture-of-experts ranker for Google Discover's heterogeneous feed. The model injects explicit heterogeneity signals h (content type, followed-creator, shoppable, AI-summary) into task-specific gating and into expert representations via FiLM-style modulation. The authors also introduce DL-AUC, a linear combination of Micro-AUC and Macro-xAUC, and LENS, an observability framework with activation slicing and permutation-invariant expert matching. On a 7-day, ~10M-example holdout, HA-MoE reports DL-AUC gains over Shared MLP and Standard MMoE, and it reduces the Web-vs-Video cross-type xAUC gap from 0.141 to 0.060 on pDisinterest. Online A/B tests on 1% of traffic report positive lifts with tight 95% CIs.

Significance. At its best, the paper offers one of the few public end-to-end accounts of heterogeneous feed ranking at industrial scale. The online A/B evidence is the most credible part: lifts in DAU, scroll depth, and diverse-feed metrics have small confidence intervals and are directionally consistent with the offline story. The reported deployment overhead (<5% model size, <0.5% latency) is also valuable. However, the causal attribution to explicit heterogeneity signals is not yet established. The ablations in Table 1 are flat on the headline pDisinterest metric, no control removes h or matches added capacity, and the DL-AUC metric is both author-defined and used as the main offline selector, with a mathematical protection claim that is not generally valid. If these points are addressed, the paper would be a solid RecSys industrial case study.

major comments (3)
  1. [§4.2, Table 1; §3.2, Eq. (3)] The central claim that 'both components' (HA-Gating and HDLM) contribute is not supported by the ablations. HA-MoE and HA-MoE w/o HA-Gating both reach pDisinterest DL-AUC 0.949; the pInterest difference (0.691 vs 0.689) is within rounding. Since 'w/o HA-Gating' still contains HDLM and 'w/o HDLM' still contains HA-Gating, h is never removed entirely. The modulation matrices in Eq. (3) add parameters, so the gain over Standard MMoE could be added capacity rather than semantic heterogeneity. Moreover, h includes several signals (followed-creator, shoppable, AI-summary), but the evaluation only segments by content type. Please add an h→random-noise control and an equal-capacity non-heterogeneous augmentation, and report per-seed variance.
  2. [§4.1.1, Eqs. (10)–(13)] The statement that the 'marginal cross-type gains... cannot mathematically offset the severe degradation' is not a mathematical fact. With a=xAUC(Web+, Vid−) and b=xAUC(Vid+, Web−), uniformly raising Web scores can increase a to at most 1 and reduce b to at least 0; the gain 1−a can exceed the loss b whenever a+b<1. Thus DL-AUC's protection is empirical, not guaranteed. Also λ=0.8 is fixed without sensitivity analysis, and because DL-AUC's Macro-xAUC directly rewards the cross-type gap HA-MoE was designed to close, the offline improvement is partly metric-circular. Please sweep λ and provide a concrete proof or a counterexample-aware caveat.
  3. [§4.2, Tables 1–2] No uncertainty quantification is reported for offline metrics. The ablation differences are 0.001–0.004 on DL-AUC, which is in the noise of a single evaluation at 10M examples (and possibly seed variation). The online A/B table includes confidence intervals, so the infrastructure exists. Please report standard errors or confidence intervals across seeds/evaluation samples and, where possible, paired significance tests. Without this, the claim that HDLM 'reverses the regression' (pDisinterest 0.949 vs 0.934) cannot be separated from chance.
minor comments (5)
  1. [§3.1/§4.2] The aggregation from 11 tasks to pInterest/pDisinterest is undefined. State whether DL-AUC is averaged unweighted over tasks and how macro-categories are formed.
  2. [§4.1.2] Several training hyperparameters (α, initial w_t, resampling ratios) are said to inherit production configs but not reported; this limits reproducibility.
  3. [§4.1.1] 'GAUC' is mentioned as the ideal but not defined; consider defining or citing it.
  4. [§3.2, Eq. (7)] The pairwise loss normalization 1/(|B+||B−|) appears to correspond to all pairs, but the text says pairs are constructed per task and filtered; clarify the effective sampling distribution.
  5. [§3.3] Consider stating explicitly that PIEM compares task-specialization behavior, not parameter-level equivalence; the current discussion is clear but a one-sentence limitation on warm-start interpretation would help.

Circularity Check

0 steps flagged

No significant circularity: HA-MoE's offline gains are measured by an explicitly defined evaluation metric and independently grounded by online A/B testing, not derived from the metric by construction.

full rationale

The paper's claimed derivation chain is not circular. HA-MoE is trained with a pointwise BCE plus pairwise RankNet loss (Eqs. 5–7), not with DL-AUC or any heterogeneity-separation objective. DL-AUC (Eq. 10) is an evaluation metric introduced for measurement, defined as a weighted blend of Micro-AUC and Macro-xAUC (Eqs. 11–13); it is not fitted to the model and does not appear in the training loss. The fact that Macro-xAUC rewards cross-type ranking correctness and HA-MoE uses content-type signals as inputs is a substantive design alignment, not a definitional equivalence; the model could in principle use h without improving xAUC, and indeed the paper reports concrete xAUC changes (Table 2). The offline results are further corroborated by live A/B tests (Table 3), providing external grounding. The paper contains no load-bearing self-citations: xAUC, MMoE, FiLM, and other cited methods are independent, prior published work. The only notable concerns are correctness/attribution issues, not circularity: the ablation in Table 1 shows HA-MoE w/o HA-Gating achieves the same pDisinterest DL-AUC (0.949) as the full HA-MoE, undermining the claim that both components contribute to the headline negative-prediction gain, and the 'cannot mathematically offset' argument in Sec. 4.1.1 is not generally true for Macro-xAUC averages. These are empirical or logical weaknesses, but they do not make the derivation equivalent to its inputs. Accordingly, the circularity score is 0.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The central claim rests on proprietary empirical evaluation rather than a mathematical derivation. Several hyperparameters are inherited production choices or author-set policy weights. HA-MoE, LENS, PIEM, and DL-AUC are methodological constructs, not new physical entities needing independent empirical handles.

free parameters (5)
  • Number of experts N = 4
    Selected from {2,4,8,16} via grid search to maximize representational power within fixed production budgets. Affects all reported results and the claimed deployment overhead.
  • DL-AUC policy weight λ = 0.8
    Hand-set 'business policy weight' balancing Micro-AUC and Macro-xAUC. The headline offline improvements depend on this choice; no sensitivity analysis is provided.
  • Loss balancing coefficient α = not reported (inherited production value)
    Balances pointwise BCE and pairwise RankNet losses in Eq. (5). Its value is not disclosed.
  • Initial task weights w_t = not reported
    Initial values for 11 task heads inherit production configuration and are then adjusted by gradient norms. Not specified.
  • Positive upsampling / negative downsampling ratios = not reported
    Training logs undergo positive upsampling and negative downsampling to mitigate class imbalance; ratios are not stated.
axioms (4)
  • domain assumption The explicit heterogeneity signal h (content_type, followed-creator, shoppable, AI-enhanced card) is a sufficient and non-redundant proxy for the structural content heterogeneity that causes negative transfer.
    Used throughout Section 3.1 and 3.2 to justify HA-Gating and HDLM. If h duplicates information already in dense features x, the attribution of gains to heterogeneity adaptation is unsupported.
  • ad hoc to paper DL-AUC with λ=0.8 and a 7-day sampled holdout is a valid proxy for production feed health and user value.
    Introduced in Section 4.1.1 by the authors; its components and the λ choice are not externally anchored, and no sensitivity analysis is given.
  • domain assumption Co-impressed positive-negative pairs within the same user session are a valid approximation of the true ranking objective for the pairwise RankNet loss.
    Invoked after Eq. (7) in Section 3.2; this is a standard industrial approximation but can bias learning toward head items or session-specific artifacts.
  • domain assumption The online A/B test on 1% of traffic over 7 days, under production guardrails, is free of novelty effects and interference that would invalidate the lifts.
    Used in Section 4.4 to translate short-term engagement lifts into a claim of ecosystem-health improvement.

pith-pipeline@v1.3.0-daily-deepseek · 12893 in / 13080 out tokens · 117722 ms · 2026-08-01T05:21:03.958653+00:00 · methodology

0 comments
read the original abstract

Heterogeneous recommendation feeds present complex challenges that extend beyond those found in highly homogeneous environments (e.g., music-only or video-only closed-ecosystem platforms). In Google Discover, a unified feed integrates diverse content sourced from the decentralized open web, including web articles, long-form and short-form videos, user-generated content (UGC), and beyond. Different content types exhibit distinct feature densities and user interaction patterns. Building a unified ranking model that sustains high performance across such heterogeneity, while avoiding negative transfer or majority bias, remains a significant industrial challenge. This paper presents an end-to-end case study on the industrial-scale multi-task ranking of heterogeneous feeds, grounded in real-world deployment. We introduce HA-MoE, a heterogeneity-adaptive multi-gated mixture-of-experts architecture that incorporates explicit heterogeneity context into both gating networks and expert representations. This approach enables effective specialization without significantly increasing operational overhead. To support reliable deployment, we introduce LENS, a lightweight observability framework that provides interpretable diagnostics of expert specialization and tracks this functional heterogeneity across continuous retraining. We evaluate our method using Dual-Level AUC (DL-AUC), a heterogeneity-aware evaluation metric that combines global ranking performance with cross-segment ranking correctness. Offline evaluations on a large-scale industrial dataset demonstrate consistent improvements over baseline models. Furthermore, online A/B testing confirms gains in feed activity and exploration metrics. Together, offline and online results validate the effectiveness of our approach for managing heterogeneity in industrial-scale recommender systems.

Figures

Figures reproduced from arXiv: 2607.27577 by Di Bai, Jintao Liu, Luoshu Wang, Nada Al-Thawr, Peifan Wu, Zhenwei Tang.

Figure 1
Figure 1. Figure 1: Overview of Google Discover. The system constructs [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of HA-MoE. Explicit heterogeneity signals [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Expert Activation for Universal Tasks. Task A and [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Expert Activation for Specialized Tasks. Tasks C and [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Task specialization tracking using task-expert acti [PITH_FULL_IMAGE:figures/full_fig_p007_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.

Reference graph

Works this paper leans on

38 extracted references · 6 linked inside Pith

  1. [1]

    Vito Walter Anelli, Alejandro Bellogín, Antonio Ferrara, Daniele Malitesta, Fe- lice Antonio Merra, Claudio Pomo, Francesco Maria Donini, Eugenio Di Sciascio, and Tommaso Di Noia. 2021. How to perform reproducible experiments in the ELLIOT recommendation framework: Data processing, model selection, and performance evaluation.IIR(2021)

    2021

  2. [2]

    Ludovico Boratto, Gianni Fenu, and Mirko Marras. 2021. Interplay between upsampling and regularization for provider fairness in recommender systems. User Modeling and User-Adapted Interaction31, 3 (2021), 421–455

    2021

  3. [3]

    Marc Brockschmidt. 2020. GNN-FiLM: Graph neural networks with feature- wise linear modulation. InInternational conference on machine learning. PMLR, 1144–1152

    2020

  4. [4]

    Chris Burges, Tal Shaked, Erin Renshaw, Ari Lazier, Matt Deeds, Nicole Hamilton, and Greg Hullender. 2005. Learning to rank using gradient descent. InProceedings of the 22nd international conference on Machine learning. 89–96

    2005

  5. [5]

    Rich Caruana. 1997. Multitask learning.Machine Learning28, 1 (1997), 41–75

    1997

  6. [6]

    Jianxin Chang, Chenbin Zhang, Yiqun Hui, Dewei Leng, Yanan Niu, Yang Song, and Kun Gai. 2023. PEPNet: Parameter and Embedding Personalized Network for Infusing with Personalized Prior Information. InProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 3795–3804

    2023

  7. [7]

    Xu Chen, Jingsen Zhang, Lei Wang, Quanyu Dai, Zhenhua Dong, Ruiming Tang, Rui Zhang, Li Chen, Xin Zhao, and Ji-Rong Wen. 2023. REASONER: an explainable recommendation dataset with comprehensive labeling ground truths.Advances in Neural Information Processing Systems36 (2023), 14497–14515

    2023

  8. [8]

    Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and Andrew Rabinovich. 2018. GradNorm: Gradient normalization for adaptive loss balancing in deep multitask networks. InInternational conference on machine learning. PMLR, 794–803

    2018

  9. [9]

    Paul Covington, Jay Adams, and Emre Sargin. 2016. Deep neural networks for youtube recommendations. InProceedings of the 10th ACM Conference on Recommender Systems. 191–198

    2016

  10. [10]

    Gideon Dror, Noam Koenigstein, Yehuda Koren, and Markus Weimer. 2012. The Yahoo! music dataset and KDD-Cup ’11. InProceedings of KDD Cup 2011. PMLR, 3–18

    2012

  11. [11]

    Michael D Ekstrand, Ben Carterette, and Fernando Diaz. 2024. Distributionally- informed recommender system evaluation.ACM Transactions on Recommender Systems2, 1 (2024), 1–27

    2024

  12. [12]

    Tom Fawcett. 2006. An introduction to ROC analysis.Pattern Recognition Letters 27, 8 (2006), 861–874

    2006

  13. [13]

    William Fedus, Jeff Dean, and Barret Zoph. 2022. A review of sparse expert models in deep learning.arXiv preprint arXiv:2209.01667(2022)

    Pith/arXiv arXiv 2022

  14. [14]

    Chongming Gao, Shijun Li, Wenqiang Lei, Jiawei Chen, Biao Li, Peng Jiang, Xiangnan He, Jiaxin Mao, and Tat-Seng Chua. 2022. KuaiRec: A fully-observed dataset and insights for evaluating recommender systems. InProceedings of the 31st ACM International Conference on Information & Knowledge Management. 540–550

    2022

  15. [15]

    Kalervo Järvelin and Jaana Kekäläinen. 2002. Cumulated gain-based evaluation of IR techniques.ACM Transactions on Information Systems (TOIS)20, 4 (2002), 422–446

    2002

  16. [16]

    Nathan Kallus and Angela Zhou. 2019. The fairness of risk scores beyond classi- fication: Bipartite ranking and the xAUC metric.Advances in neural information processing systems32 (2019)

    2019

  17. [17]

    Harold W. Kuhn. 1955. The Hungarian method for the assignment problem. Naval Research Logistics Quarterly2, 1-2 (1955), 83–97

    1955

  18. [18]

    Ningyuan Li, Yunxuan Ma, Yang Zhao, Qian Wang, Zhilin Zhang, Chuan Yu, Jian Xu, Bo Zheng, and Xiaotie Deng. 2024. Ad vs organic: Revisiting incentive compatible mechanism design in e-commerce platforms. InProceedings of the ACM Web Conference 2024. 235–244

    2024

  19. [19]

    Tianqiao Liu, Zhiwei Wang, Jiliang Tang, Songfan Yang, Gale Yan Huang, and Zitao Liu. 2019. Recommender systems with heterogeneous side information. In The World Wide Web Conference. 3027–3033

    2019

  20. [20]

    Weiwen Liu and Robin Burke. 2018. Personalizing fairness-aware re-ranking. arXiv preprint arXiv:1809.02921(2018)

    Pith/arXiv arXiv 2018

  21. [21]

    Zongtao Liu, Bin Ma, Quan Liu, Jian Xu, and Bo Zheng. 2021. Heterogeneous graph neural networks for large-scale bid keyword matching. InProceedings of the 30th ACM International Conference on Information & Knowledge Management. 3976–3985

    2021

  22. [22]

    Chengqiang Lu, Mingyang Yin, Shuheng Shen, Luo Ji, Qi Liu, and Hongxia Yang. 2022. Deep unified representation for heterogeneous recommendation. In Proceedings of the ACM Web Conference 2022. 2141–2152

    2022

  23. [23]

    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. InProceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining. 1930–1939

    2018

  24. [24]

    Changhua Pei, Yi Zhang, Yongfeng Zhang, Fei Sun, Xiao Lin, Hanxiao Sun, Jian Wu, Peng Jiang, Junfeng Ge, Wenwu Ou, et al. 2019. Personalized re-ranking for recommendation. InProceedings of the 13th ACM conference on recommender systems. 3–11

    2019

  25. [25]

    Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. 2018. FiLM: Visual reasoning with a general conditioning layer. In Proceedings of the AAAI conference on artificial intelligence, Vol. 32

    2018

  26. [26]

    Shaan Shah and Meenakshi Khosla. 2025. Representational Alignment Across Model Layers and Brain Regions with Multi-Level Optimal Transport.arXiv preprint arXiv:2510.01706(2025)

    Pith/arXiv arXiv 2025

  27. [27]

    Xiang-Rong Sheng, Liqin Zhao, Guorui Zhou, Xinyao Ding, Binding Dai, Qiang Luo, Siran Yang, Jingshan Lv, Chi Zhang, Hongbo Deng, et al. 2021. One model to serve all: STAR topology adaptive recommender for multi-domain CTR pre- diction. InProceedings of the 30th ACM International Conference on Information & Knowledge Management. 4104–4113

    2021

  28. [28]

    Anke Tang, Li Shen, Yong Luo, Nan Yin, Lefei Zhang, and Dacheng Tao. 2024. Merging multi-task models via weight-ensembling mixture of experts.arXiv preprint arXiv:2402.00433(2024)

    Pith/arXiv arXiv 2024

  29. [29]

    Hongyan Tang, Junning Liu, Ming Zhao, and Xudong Gong. 2020. Progressive layered extraction (PLE): A novel multi-task learning (MTL) model for personal- ized recommendations. InProceedings of the 14th ACM conference on recommender systems. 269–278

    2020

  30. [30]

    Yue Wang, Dawei Yin, Luo Jie, Pengyuan Wang, Makoto Yamada, Yi Chang, and Qiaozhu Mei. 2016. Beyond ranking: Optimizing whole-page presentation. In Proceedings of the Ninth ACM International Conference on Web Search and Data Mining. 103–112

    2016

  31. [31]

    Mingyang Wu, Li Lin, Wenbin Zhang, Xin Wang, Zhenhuan Yang, and Shu Hu

  32. [32]

    Ruobing Xie, Shaoliang Zhang, Rui Wang, Feng Xia, and Leyu Lin. 2021. Hierar- chical reinforcement learning for integrated recommendation. InProceedings of the AAAI conference on artificial intelligence, Vol. 35. 4521–4528

    2021

  33. [33]

    Guanghu Yuan, Fajie Yuan, Yudong Li, Beibei Kong, Shujie Li, Lei Chen, Min Yang, Chenyun Yu, Bo Hu, Zang Li, et al . 2022. Tenrec: A large-scale multipurpose benchmark dataset for recommender systems.Advances in Neural Information Processing Systems35 (2022), 11480–11493

    2022

  34. [34]

    Jun Yuan, Guohao Cai, and Zhenghua Dong. 2024. A parameter update balancing algorithm for multi-task ranking models in recommendation systems. In2024 IEEE International Conference on Data Mining (ICDM). IEEE, 929–934

    2024

  35. [35]

    Zijian Zhang, Shuchang Liu, Jiaao Yu, Qingpeng Cai, Xiangyu Zhao, Chunxu Zhang, Ziru Liu, Qidong Liu, Hongwei Zhao, Lantao Hu, et al . 2024. M 3oE: Multi-domain multi-task mixture-of-experts recommendation framework. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 893–902

    2024

  36. [36]

    Xiangyu Zhao, Changsheng Gu, Haoshenglun Zhang, Xiwang Yang, Xiaobing Liu, Jiliang Tang, and Hui Liu. 2021. DEAR: Deep reinforcement learning for online advertising impression in recommender systems. InProceedings of the AAAI conference on artificial intelligence, Vol. 35. 750–758

    2021

  37. [37]

    Yuang Zhao, Zhaocheng Du, Qinglin Jia, Linxuan Zhang, Zhenhua Dong, and Ruiming Tang. 2024. Retrievable Domain-Sensitive Feature Memory for Multi- Domain Recommendation.arXiv preprint arXiv:2405.12892(2024)

    Pith/arXiv arXiv 2024

  38. [2025]

    Preserving AUC fairness in learning with noisy protected groups.arXiv preprint arXiv:2505.18532(2025)

    Pith/arXiv arXiv 2025