Pith. sign in

REVIEW 5 major objections 4 minor 28 references

An Audio-centric Multi-task Learning Framework for Streaming Ads Targeting on Spotify

T0 review · 5 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that modality-aware task grouping inside a mixture-of-experts model—splitting click prediction into audio and video heads, masking losses by modality, and using DCN-v2 experts—is the near-optimal configuration for…

desk verdict Useful deployed MTL recipe with a genuine A/B test, but the headline 14.5% CTR gain is an uncontrolled rollout number mislabeled as an A/B result. read the letter →

arxiv 2506.18735 v1 pith:MLZJQSCL submitted 2025-06-23 cs.IR eess.AS

Shivam Verma , Vivian Chen , Darren Mei This is my paper
classification cs.IReess.AS
keywords computationaladvertisingclick-throughratepredictionmulti-tasklearningmixture-of-expertsaudio-centricplatformdeepandcrossnetworkadaptivelossmaskingonlineA/Btest
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

The paper tries to establish that click-through prediction for ads on an audio-first platform has a structural problem—over 70% of listening happens out of focus, audio impressions dwarf video, and out-of-focus impressions click at roughly a tenth of the rate—and that a multi-task model must be organized around that asymmetry. It proposes CAMoE, a mixture-of-experts network whose two task heads are split by ad modality rather than by content or by individual slot, with adaptive loss masking so the dominant audio task cannot drown out video learning, and DCN-v2 blocks as experts to capture feature interactions. Offline, the two-head configuration improves AUC-PR across all ad slots relative to Spotify's Wide & Deep production baseline and is presented as near-Pareto-optimal against one-task, seven-task, and content-split alternatives. Online, a controlled A/B test shows 10% audio and 1.4% video CTR gains, and post-rollout figures reach 14.5% and 1.3% with lower cost-per-click. If correct, this gives other audio-centric or backgrounded platforms a concrete recipe for ad targeting rather than treating video-style models as the default.

What carries the argument

The load-bearing object is CAMoE, the Cross-modal Adaptive Mixture-of-Experts architecture: a modified Multi-gate Mixture-of-Experts (MMoE) network in which the click prediction task is divided into two modality-aware heads, one for audio (plus its display leave-behind slots) and one for video. Three components carry the argument. Modality-aware task grouping lets each head specialize in the engagement patterns of a modality. Adaptive loss masking restricts each task's loss to examples from its own modality, preventing the abundant audio data from overwhelming video learning. DCN-v2 blocks inside each expert capture high-order feature interactions, and per-task temperature scaling calibrates the final probabilities for auction bidding. Together these components are what the paper credits for the AUC-PR, calibration, and online CTR gains.

What would settle it

Re-run the budget-split A/B test that the paper describes, serving CAMoE and the old Wide & Deep model at the same time to identical eligible audio and video impressions; if the audio CTR lift falls short of the 10% reported in the controlled test, or disappears when DCN experts or adaptive loss masking are removed, the central causal claim is not supported.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is that CAMoE—a two-head multi-task network built on a mixture-of-experts base—is the right architecture for click-through prediction on an audio-centric, multi-modal ad platform. The audio head covers audio slots and the display leave-behind slots that ride on them; the video head covers in-stream and embedded video slots. Each head's loss is masked so that only same-modality examples update it, and each of the two experts is a DCN-v2 block that explicitly models feature crosses. In offline evaluation against Spotify's Wide & Deep production baseline, this configuration raises AUC-PR for every ad slot—for example 20.73% for Stream Video, 24.10% for Stream Audio, and 43.27% for Podcast Video—and cuts expected calibration error, while content-based and seven-slot task splits do not match it. In a budget-split online A/B test, CAMoE improved audio CTR by 10% and video CTR by 1.4%, and after full rollout the paper reports 14.5% higher audio CTR, 1.3% higher video CTR, and 4.8% lower eCPC on audio slots.

Load-bearing premise

The load-bearing premise is that the post-deployment performance gains—14.5% higher audio CTR and 4.8% lower audio eCPC—are caused by CAMoE, even though those figures were recorded after a full rollout without a simultaneous randomized control that rules out concurrent changes to campaigns, inventory, pacing, or the auction.

Editorial extensions

If this is right

  • Splitting the click task into two modality heads outperforms both a single all-slot model and a separate head per ad slot, especially on the high-value Stream Audio and Stream Video slots.
  • Adaptive loss masking is a load-bearing component: without it, video-slot AUC-PR turns negative even though audio performance stays positive, showing the audio majority would otherwise dominate shared learning.
  • Using DCN-v2 blocks as MoE experts is what makes the two-task setup beat the single-task setup; removing DCN flips the comparison.
  • Modality-based grouping beats content-based (music vs. podcast) grouping, so the relevant axis for this platform is how the user attends to the ad rather than what content it accompanies.
  • Deployed at scale, the model yields the stated CTR and cost-per-click improvements, implying that the offline AUC-PR and ECE gains translate into auction outcomes.

Reading between the lines

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

  • A testable extension the paper leaves implicit: the same modality-split recipe should be evaluated on other audio-first or backgrounded services (podcast platforms, navigation apps, smart-speaker assistants), where a dominant low-attention audio modality should create the same imbalance problem; if the gains replicate, the mechanism is attention modality rather than Spotify-specific inventory.
  • The expert-masking experiment implies that further video-side gains are on the table if interference from the audio-leaning expert can be removed cheaply; the paper rejected naive masking on latency grounds, so a sparse or early-exit routing variant is a concrete next test.
  • One consequence the authors do not draw: the controlled A/B test (+10% audio, +1.4% video) is the cleaner causal estimate, while the 100% rollout deltas (+14.5%, +1.3%, -4.8% eCPC) mix CAMoE's effect with whatever else changed at deployment, so a synthetic-control analysis of the rollout time series would sharpen the causal number.
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

5 major / 4 minor

Summary. The paper presents CAMoE, a multi-task learning framework for click-through rate (CTR) prediction on Spotify's audio-centric advertising platform. The key design choices are modality-aware task grouping (audio vs. video heads), adaptive loss masking to mitigate audio-dominated data imbalance, and DCN-v2 blocks as Mixture-of-Experts components. The authors report offline AUC-PR gains over a Wide & Deep production baseline, ablation studies for task grouping, loss masking, and DCN experts, and an online evaluation that combines a budget-split A/B test with post-deployment 100% rollout metrics. The central claim, stated in the Abstract and Conclusion, is that deployment of CAMoE yielded a 14.5% CTR increase for audio ads, a 1.3% increase for video ads, and a 4.8% eCPC reduction for audio slots, and that the 2-task configuration is near-Pareto-optimal across ad formats.

Significance. If the claims are fully supported, CAMoE would be a valuable industrial contribution: it is a deployed system, it addresses a genuine modality-imbalance problem that is under-studied in the recommender-systems literature, and its controlled A/B test provides an external benchmark that is not used to fit hyperparameters. The paper also gives credit for including a large-scale deployment, detailed ablations, a calibration analysis, and an appendix study of inference-time expert masking. However, the significance is materially reduced by the evidence gaps: the headline online numbers come from an uncontrolled rollout rather than the randomized comparison, several offline tables are internally inconsistent, and the ablations do not cleanly isolate the contributions of the proposed components. These issues must be resolved before the central claims can be taken at face value.

major comments (5)
  1. [Abstract; §6.3, Table 6; §7] The Abstract and §7 state that 'in an online A/B test, CAMoE improved CTR by 14.5% for audio ads and 1.3% for video ads,' but Table 6 reports A/B test results of +10% for audio CTR and +1.4% for video CTR. The 14.5%, 1.3%, and eCPC reductions are instead 100% rollout results, i.e., observational before/after comparisons, not a randomized experiment. This distinction matters because Eq. (1) shows Final Bid depends on pCTR and the 24-hour average CTR b, so deploying a new pCTR model changes b and hence the set of winning ads and measured CTR/eCPC even if ranking quality is unchanged; concurrent changes to campaigns, inventory, pacing, or auction logic are not controlled. The abstract's central assertion therefore exceeds the causal evidence.
  2. [§6.2.1, Tables 2 and 3] The offline results are internally inconsistent. The text in §6.2.1 says the 2-task CAMoE improves Stream Video by 18.03%, Stream Audio by 16.10%, and Podcast by 79.62%, but Table 2's CAMoE(t=2) row reports 20.73%, 24.10%, and 54.21% for those same slots. Table 3's Audio-Video row, which is supposed to be the same 2-task modality grouping, reports yet another set of values (17.69%, 18.78%, and 37.10% for Stream Video, Stream Audio, and Podcast). Because the paper's central offline claim of near-Pareto-optimal performance relies on these exact numbers, the discrepancies must be resolved and the tables made mutually consistent.
  3. [§6.2.3–§6.2.4, Tables 4 and 5] The ablations do not separate the contributions of adaptive loss masking (ALM) and DCN experts. The 'Without DCN' row in Table 5 is numerically identical to the MMoE(t=2) row in Table 2, so removing DCN also removes any accompanying CAMoE-specific modifications to gating, loss masking, or task grouping; it is not a controlled removal of DCN from CAMoE. Similarly, Table 4's 'No Masking' row has no matching baseline configuration in Table 2, leaving unclear what else may have changed. The claims that ALM and DCN each provide large, significant gains are therefore not supported by the presented controlled comparisons.
  4. [§6.3, Table 6] The controlled A/B test is under-specified. The paper does not state the randomization unit (user, session, or market), the number of observations, confidence intervals, standard errors, or whether standard errors account for within-market correlation. A budget-split design can also push the control model toward harder inventory if budgets are reallocated, which would inflate the measured lift even without a model-quality improvement. As reported, the +10% and +1.4% CTR effects cannot be fully interpreted as unbiased causal estimates of the model's ranking improvement.
  5. [§6.2, Tables 2–5] All offline results are reported only as percentage changes relative to the Wide & Deep baseline; no absolute AUC-PR values, standard deviations, or sample sizes are given. This makes it impossible to assess the practical magnitude of the reported gains or to check whether the differences among configurations are meaningful beyond the stated p-value thresholds.
minor comments (4)
  1. [§6.2.4] There is a typo in the sentence 'Without DCN experts, the 2-task CAMoE performs worse than the the respective single-task setup'; 'the the' should be 'the'.
  2. [Eq. (2)] The notation y_i^(n) and \hat{y}_i^(m) is not defined explicitly; please clarify that the superscript indexes the modality of the label and the prediction, and that the indicator I{n=m} selects the label modality.
  3. [§A.1, Fig. 7] The Pareto analysis plots only Stream Audio and Stream Video, while §A.1.1 claims to consider all ad slots, calibration, and online metrics; the appendix should state more explicitly that Fig. 7 is a two-dimensional projection and that the broader Pareto claim is qualitative.
  4. [§A.2.5, Table 7] The expert-masking results report large relative changes (e.g., +122.4% for Embedded Music) without confidence intervals or test-set sizes; given the small volume of those slots, some of these effects may be within noise.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the paper's central results are measured outcomes from offline comparisons and an online A/B test, not quantities derived from fitted inputs.

full rationale

The paper's central claims are empirical measurements rather than derived predictions: Table 6 reports A/B test and rollout deltas for CTR, eCPC, and clicks, and the offline sections report AUC-PR and ECE changes relative to a Wide & Deep production baseline. None of these quantities is constructed from the model's own parameters in a way that forces the reported result. The architecture components (MMoE, DCN-v2, temperature scaling) are supported by external, independently published prior work, and the paper contains no load-bearing self-citation chain. The 'near-Pareto optimal' judgment in Appendix A.1 is explicitly framed relative to the authors' chosen priority slots (Stream Audio and Stream Video) and is stated as a balancing argument across tables, not an imported uniqueness theorem or a definitional equivalence. The apparent mismatch between the A/B test figures (+10% audio CTR, +1.4% video CTR) and the 100% rollout figures (+14.5%, +1.3%), as well as the causal attribution of rollout deltas given the feedback loop in Eq. (1), are concerns about experimental validity and causal identifiability, not about circularity in the derivation chain. No fitted parameter is renamed as a prediction, and no equation reduces to its own input by construction.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claims rest on private evaluation data, the chosen metric set, the focus-state model, and the audio/video grouping premise. No new physical or mathematical entities are introduced.

free parameters (4)
  • lambda_m (modality loss weights) = not reported (tuned)
    Eq. 3 scales each modality's masked loss, with lambda_audio + lambda_video = 1. The value governs the audio/video balance and is set by hyperparameter search, but the chosen values are not disclosed.
  • Temperature T per task = not reported (tuned)
    Eq. 5 temperature scaling for post-training calibration; one T is tuned per task. The values are not listed.
  • DCN architecture hyperparameters = not reported
    Section 5.2: number of DCN-v2 cross layers L, bottleneck rank r, number of experts (K=2), and tower hidden sizes are chosen by the authors but not specified.
  • Downsampling ratio for majority slots = not reported
    Section 4.2 states that down-sampling plus adaptive loss masking was used, but the downsampling ratio is not given.
assumptions (5)
  • domain assumption The in-focus/out-of-focus focus-state model and the tenfold CTR disparity between states are accurate.
    Section 3.1 cites 'internal analysis' for over 70% out-of-focus listening and a 10x CTR difference; the entire audio/video task split is motivated by this premise.
  • domain assumption AUC-PR is the appropriate primary offline metric for comparing models under extreme class imbalance.
    Section 6.1 justifies AUC-PR over AUC-ROC and accuracy because of imbalance; the paper does not report AUC-ROC or logloss, so conclusions depend on this metric choice.
  • ad hoc to paper Display leavebehind slots should be grouped with the audio task because they are not counted as additional impressions.
    Section 4.1 makes this grouping decision, which affects what the audio head learns, but the grouping is not independently validated.
  • standard math Standard DCNv2 cross-layer update (Eq. 4) and MMoE gating (Eq. 8) are valid building blocks that transfer to this ad setting without modification.
    Eqs. 4 and 8 are taken from prior work (Wang et al., Ma et al.), and no derivation or adaptation is provided in this paper.
  • domain assumption The auction formula Final Bid = min(o / (1 + p * c * b), o) correctly models the serving environment.
    Section 3.2, Eq. 1; if the actual auction differs, the relationship between pCTR, revenue, and the reported CTR/eCPC metrics could change.

how reviews work

0 comments
Cite this review

Pith. "Pith review of An Audio-centric Multi-task Learning Framework for Streaming Ads Targeting on Spotify." pith.science (2026). https://pith.science/paper/MLZJQSCL

@misc{pith2026250618735,
  author       = {Pith},
  title        = {Pith review of: An Audio-centric Multi-task Learning Framework for Streaming Ads Targeting on Spotify},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MLZJQSCL}},
  note         = {Machine review of arXiv:2506.18735}
}
read the original abstract

Spotify, a large-scale multimedia platform, attracts over 675 million monthly active users who collectively consume millions of hours of music, podcasts, audiobooks, and video content. This diverse content consumption pattern introduces unique challenges for computational advertising, which must effectively integrate a variety of ad modalities, including audio, video, and display, within a single user experience. Traditional ad recommendation models, primarily designed for foregrounded experiences, often struggle to reconcile the platform's inherent audio-centrality with the demands of optimizing ad performance across multiple formats and modalities. To overcome these challenges, we introduce Cross-modal Adaptive Mixture-of-Experts (CAMoE), a novel framework for optimizing click-through rate (CTR) prediction in both audio-centric and multi-modal settings. CAMoE enhances traditional mixture-of-experts models by incorporating modality-aware task grouping, adaptive loss masking, and deep-cross networks (DCN) to capture complex feature interactions within a multi-modal ad ecosystem. Through extensive ablation studies, we demonstrate that this approach achieves near Pareto-optimal performance across audio, video, and display ad formats, significantly improving AUC-PR compared to conventional single-task and content-based multi-task learning baselines. When deployed at scale on Spotify's ad serving platform, CAMoE delivered substantial gains, yielding a 14.5% increase in CTR for audio ads, a 1.3% increase for video ads, and a 4.8% reduction in expected cost-per-click (eCPC) for audio slots.

Figures

Figures reproduced from arXiv: 2506.18735 by the authors.

Figure 1
Figure 1. Left to right: (a) An audio ad stream. (b) An unmuted [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. High-level overview of the ads ranking system at [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. (A) Independent uni-modal pClick models, for each ad surface or modality (B) Single-task pClick model incorporating [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Cross-modal Adaptive Mixture-of-Experts (CAMoE) [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Calibration plots for the Stream Audio and Stream [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Calibration plots for the Stream Audio and Stream [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Analysis of Pareto Optimality for various CAMoE, [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 12 canonical work pages

  1. [1]

    Fedor Borisyuk, Mingzhou Zhou, Qingquan Song, Siyu Zhu, Birjodh Tiwana, Ganesh Parameswaran, Siddharth Dangi, Lars Hertel, Qiang Charles Xiao, Xi- aochen Hou, et al . 2024. LiRank: Industrial Large Scale Ranking Models at LinkedIn. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 4804–4815

    work page 2024

  2. [2]

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

    work page 1997

  3. [3]

    Heng-Tze Cheng, Levent Koc, Jeremiah Harmsen, Tal Shaked, Tushar Chandra, Hrishi Aradhye, Glen Anderson, Greg Corrado, Wei Chai, Mustafa Ispir, Rohan Anil, Zakaria Haque, Lichan Hong, Vihan Jain, Xiaobing Liu, and Hemal Shah

  4. [4]

    Paul Covington, Jay Adams, and Emre Sargin. 2016. Deep Neural Networks for YouTube Recommendations. InProceedings of the 10th ACM Conference on Recommender Systems. IEEE, Boston, MA, USA. doi:10.1145/2959100.2959190

    arXiv 2016

  5. [5]

    Weinberger

    Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. 2017. On calibration of modern neural networks. InProceedings of the 34th International Conference on Machine Learning - Volume 70(Sydney, NSW, Australia)(ICML’17). JMLR.org, 1321–1330

    work page 2017

  6. [6]

    Honghao Li, Yiwen Zhang, Yi Zhang, Hanwei Li, and Lei Sang. 2024. DCNv3: Towards Next Generation Deep Cross Network for CTR Prediction.CoRR(2024)

    work page 2024

  7. [7]

    Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollár. 2017. Focal Loss for Dense Object Detection. In2017 IEEE International Conference on Computer Vision (ICCV). 2999–3007. doi:10.1109/ICCV.2017.324

    doi:10.1109/iccv.2017.324 2017

  8. [8]

    Xi Lin, Hui-Ling Zhen, Zhenhua Li, Qingfu Zhang, and Sam Kwong. 2019. Pareto Multi-Task Learning. InProceedings of the 33rd International Conference on Neural Information Processing Systems. ACM, Vancouver, Canada, 12060–12070. doi:10. 5555/3454287.3455367

    arXiv 2019

Show all 28 references
  1. [9]

    Xi Victoria Lin, Akshat Shrivastava, Liang Luo, Srinivasan Iyer, Mike Lewis, Gargi Ghosh, Luke Zettlemoyer, and Armen Aghajanyan. 2024. MoMa: Ef- ficient Early-Fusion Pre-training with Mixture of Modality-Aware Experts. arXiv:2407.21770 [cs.AI] https://arxiv.org/abs/2407.21770

    2024 arXiv

  2. [10]

    Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Philip S. Yu. 2017. Learning Multiple Tasks with Multilinear Relationship Networks. InProceedings of the 31st International Conference on Neural Information Processing Systems. ACM, Long Beach, CA, USA, 1593–1602. doi:10.5555/329...

    2017

  3. [11]

    Jiaqi Ma, Zhe Zhao, Xinyang Yi, Jilin Chen, Lichan Hong, and Ed 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 and Data Mining. ACM, London,...

    2018

  4. [12]

    Ning Ma, Mustafa Ispir, Yuan Li, Yongpeng Yang, Zhe Chen, Derek Zhiyuan Cheng, Lan Nie, and Kishor Barman. 2022. An Online Multi-task Learning Framework for Google Feed Ads Auction Models. InProceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining(W...

    2022

  5. [13]

    Xiao Ma, Liqin Zhao, Guan Huang, Zhi Wang, Zelin Hu, Xiaoqiang Zhu, and Kun Gai. 2018. Entire Space Multi-Task Model: An Effective Approach for Estimating Post-Click Conversion Rate. InProceedings of the 41st International ACM SIGIR Conference on Research & Development in Info...

    2018

  6. [14]

    Brendan McMahan, Gary Holt, D

    H. Brendan McMahan, Gary Holt, D. Sculley, Michael Young, Dietmar Ebner, Julian Grady, Lan Nie, Todd Phillips, Eugene Davydov, Daniel Golovin, Sharat Chikkerur, Dan Liu, Martin Wattenberg, Arnar Mar Hrafnkelsson, Tom Boulos, and Jeremy Kubica. 2013. Ad Click Prediction: A View...

    2013

  7. [15]

    Ozan Sener and Vladlen Koltun. 2018. Multi-Task Learning as Multi-Objective Optimization. InProceedings of the 32nd International Conference on Neural In- formation Processing Systems. ACM, Montréal, Canada, 525–536. doi:10.5555/ 3326943.3326992

    2018

  8. [16]

    Noam Shazeer, *Azalia Mirhoseini, *Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. 2017. Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer. InInternational Conference on Learning Representations. https://openreview.net/foru...

    2017

  9. [17]

    Zamir, Dawn Chen, Leonidas Guibas, Jitendra Malik, and Silvio Savarese

    Trevor Standley, Amir R. Zamir, Dawn Chen, Leonidas Guibas, Jitendra Malik, and Silvio Savarese. 2020. Which Tasks Should Be Learned Together in Multi- Task Learning?. InProceedings of the 37th International Conference on Machine Learning (ICML ’20)

    2020

  10. [18]

    Chun How Tan, Austin Chan, Malay Haldar, Jie Tang, Xin Liu, Mustafa Abdool, Huiji Gao, Liwei He, and Sanjeev Katariya. 2023. Optimizing Airbnb Search Journey with Multi-task Learning. InProceedings of the 29th ACM SIGKDD Con- ference on Knowledge Discovery and Data Mining(Long...

    2023

  11. [19]

    Ernest Wang. 2020. How We Use AutoML, Multi-Task Learning and Multi-Tower Models for Pinterest Ads. https://medium.com/pinterest-engineering/how- weuse-automl-multi-task-learning-and-multi-tower-models-for-pinterest- adsdb966c3dc99e Medium article

    2020

  12. [20]

    Ruoxi Wang, Bin Fu, Gang Fu, and Mingliang Wang. 2017. Deep & Cross Network for Ad Click Predictions. InProceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. ACM, Halifax, NS, Canada, 1487–1495. doi:10.1145/3124749.3124754

    2017

  13. [21]

    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. InProceedings of the Web Conference

    2021

  14. [22]

    Xuewei Wang, Qiang Jin, Shengyu Huang, Min Zhang, Xi Liu, Zhengli Zhao, Yukun Chen, Zhengyu Zhang, Jiyan Yang, Ellie Wen, et al. 2023. Towards the Better Ranking Consistency: A Multi-task Learning Framework for Early Stage Ads Ranking. InProceedings of AdKDD ’23

    2023

  15. [23]

    Yinwei Wei, Xiang Wang, Liqiang Nie, Xiangnan He, Richang Hong, and Tat-Seng Chua. 2019. MMGCN: Multi-modal Graph Convolution Network for Personalized Recommendation of Micro-video. InProceedings of the 27th ACM International Conference on Multimedia(Nice, France)(MM ’19). Ass...

    2019

  16. [24]

    Huaizheng Zhang, Yong Luo, Qiming Ai, Yonggang Wen, and Han Hu. 2020. Look, read and feel: Benchmarking ads understanding with multimodal multitask learning. InProceedings of the 28th ACM international conference on multimedia. 430–438

    2020

  17. [25]

    Zhe Zhao, Lichan Hong, Li Wei, Jilin Chen, Aniruddh Nath, Shawn Andrews, Aditee Kumthekar, Maheswaran Sathiamoorthy, Xinyang Yi, and Ed Chi. 2019. Recommending What Video to Watch Next: A Multitask Ranking System. InPro- ceedings of the 13th ACM Conference on Recommender Syste...

    2019

  18. [1939]

    doi:10.1145/3219819.3220007

  19. [2016]

    InProceedings of the 1st Workshop on Deep Learning for Recommender Systems(Boston, MA, USA) (DLRS 2016)

    Wide & Deep Learning for Recommender Systems. InProceedings of the 1st Workshop on Deep Learning for Recommender Systems(Boston, MA, USA) (DLRS 2016). Association for Computing Machinery, New York, NY, USA, 7–10. doi:10.1145/2988450.2988454

    2016

  20. [2021]

    doi:10.1145/3442381.3450078

    ACM, Ljubljana, Slovenia, 1785–1797. doi:10.1145/3442381.3450078

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.