Pith. sign in

REVIEW 4 major objections 5 minor 44 references

CDC: Causal Domain Clustering for Multi-Domain Recommendation

T0 review · 4 major / 5 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 When a platform serves dozens of recommendation domains, grouping them by measured causal transfer effects and optimizing each cluster's training source set separately yields consistent gains, including a 4.9% online ad-revenue lift.

desk verdict A genuinely useful domain-clustering method for multi-domain recommendation, with an honest but unspecified heuristic at the center and industrial results that need more statistical support. read the letter →

arxiv 2507.06877 v1 pith:DQ5S72M6 submitted 2025-07-09 cs.IR

Huishi Luo , Yiqing Wu , Yiwen Chen , Fuzhen Zhuang , Deqing Wang This is my paper
classification cs.IR
keywords Multi-DomainRecommendationDomainClusteringGroupingCausalDiscoveryNegativeTransferLookaheadAffinityLearning
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

Multi-domain recommendation systems that train dozens of domains in a single model often suffer from negative transfer between weakly related domains, while grouping by business logic or data similarity misses the transfer relationships that only become visible after training. This paper proposes Causal Domain Clustering (CDC), which measures transfer effects directly: one affinity matrix records how much training on one domain helps or hurts another in isolation, and a second records the same under joint training, where domains can synergize or interfere. Causal discovery is used to blend the two views through a cohesion-based coefficient, and a Co-Optimized Dynamic Clustering algorithm simultaneously searches for target clusters and for the training source set that best serves each cluster, allowing the two to differ. On two public benchmarks with 25 and 50 domains, CDC reports DomainAUC gains of 4.14‰ and 6.76‰ over the best baselines, larger gains when each cluster trains its own model, and a 4.9% eCPM lift in a 14-day online A/B test across 64 domains. If these results hold, the paper establishes that transfer-aware, source-set-separating clustering is a practical, model-agnostic pre-training step for scaling multi-domain recommendation.

What carries the argument

The carrying mechanism is a pipeline of four linked objects. First, lookahead affinity measurement: the model saves its parameters, takes a one-batch gradient update on a source domain (or a mixed set), reloads the saved state, and records the loss ratio on a target domain; a ratio below one means the update helped. Second, two affinity matrices: $M_I$ stores these ratios for isolated single-domain updates, while $M_H$ stores them for updates on a mixed set with and without the source domain, isolating interaction effects, both smoothed by a factor $\alpha$. Third, causal-distance weighting: random domain subsets act as treatments, giving a treatment-effect matrix $M_T$; the dependence contribution kernel $\kappa$ measures dependence between the columns of $M_T$ for two domains, and the causal distance $M_C[u,v]$ is the arccosine of that kernel value. The cohesion of a candidate source set is the average pairwise causal distance within it, and the interaction coefficient $\lambda_{d_v|S}\in[0,1]$ is the capped ratio of $d_v$'s average distance to the set over twice the cohesion; it blends the two views in the transfer gain $J(d_u\to d_v|S)=(1-\lambda_{d_v|S})M_I[u,v]+\lambda_{d_v|S}M_H[u,v]$. Fourth, CODC is a greedy loop that grows each training set by the largest transfer gain plus a decaying affiliation score, then reassigns each target domain to the cluster whose source set gives the highest gain, alternating until convergence.

What would settle it

Build a small synthetic multi-domain problem whose isolated and joint-training transfer effects are known by construction, so the optimal clusters and source sets can be found by exhaustive search; if CDC's output does not match or beat that optimum, or if replacing the causal-distance coefficient with a fixed constant changes nothing, then the causal weighting is not carrying the claimed work.

Watch

Extended reading notes

Core claim

CDC's central claim is that the optimal grouping of many recommendation domains is determined by measured transfer gains, not static distribution similarity, and that the best training set for a cluster is itself a decision variable: it may include domains outside the cluster. The method builds an Isolated Domain Affinity Matrix $M_I$, whose entry $M_I[u,v]$ is the smoothed ratio of domain $v$'s loss after a one-step gradient update on domain $u$ alone, and a Hybrid Domain Affinity Matrix $M_H$, computed after updating on a mixed domain set with and without $u$, which isolates synergy or interference between domains. Random domain-set samples are treated as interventions, producing a treatment-effect matrix $M_T$; the causal distance $M_C[u,v]$ is the arccosine of the dependence contribution kernel between the columns of $M_T$ for $u$ and $v$. The cohesion of a candidate source set is the average causal distance within it, and the interaction coefficient $\lambda_{d_v|S}\in[0,1]$ interpolates between $M_I$ and $M_H$ in the transfer gain $J(d_u\to d_v|S)=(1-\lambda_{d_v|S})M_I[u,v]+\lambda_{d_v|S}M_H[u,v]$. The CODC algorithm alternates between growing each training set by the largest transfer gain (plus a decaying affiliation score) and reassigning target domains to the cluster whose source set benefits them most, with a feedback loop that keeps refining $M_H$ as clusters change. On the Amazon and AliCCP datasets, CDC reports 4.14‰ and 6.76‰ DomainAUC gains over the best baseline; a split variant with an independent model per cluster reports 8.46‰ and 10.15‰; and a 14-day online test over 64 domains reports a 4.9% eCPM lift.

Load-bearing premise

The method's load-bearing premise is a guessed averaging rule: when the causal distances among the domains in a training set are large, trust the joint-training transfer measurements more, and when they are small, trust the one-on-one transfer measurements more—and this rule, capped at one, is assumed to be correct for every possible training set.

Editorial extensions

If this is right

  • Existing multi-domain recommenders can be dropped into the clusters CDC produces; the paper shows CDC-derived groupings improve both STAR and PEPNet on industrial data, so the clustering step is model-agnostic.
  • Because training source sets are optimized separately from target clusters, a cluster of sparse minor domains can be trained partly on high-volume external domains; the CDC-$S_k$ offline results indicate this is a main source of gain.
  • The deployed configuration adds almost no compute: it matches the baseline's storage and uses 1.09x the single-model training time on the industrial dataset, so the accuracy gains do not come from a heavier model.
  • The split configuration, one independent model per cluster, produces the largest offline gains (8.46‰ and 10.15‰ DomainAUC), suggesting that preventing cross-cluster interference can be worth more than sharing parameters across all domains.

Reading between the lines

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

  • A direct extension is to apply the same dual-affinity, causal-distance clustering to other transfer-learning settings where isolated and joint behavior differ, such as grouping tasks in computer vision or grouping clients in personalized federated learning.
  • Because the online A/B test deployed only the unified-model version, the larger offline gains of the per-cluster split version remain untested live; running that configuration online is a natural next experiment.
  • The interaction coefficient in Eq. (10) is introduced heuristically rather than derived from the causal framework, so varying its functional form (for example, replacing the average-distance ratio with a learned or likelihood-based weight) is a direct way to test whether the causal-distance integration, rather than the overall dual-matrix design, causes the gains.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes Causal Domain Clustering (CDC), a framework for clustering large numbers of recommendation domains before multi-domain model training. CDC builds two inter-domain affinity matrices from one-step lookahead loss changes: an Isolated matrix M_I for pairwise transfer under independent training, and a Hybrid matrix M_H for transfer under joint training. These are combined into a transfer gain J using a cohesion-based coefficient lambda computed from a causal distance matrix M_C. A Co-Optimized Dynamic Clustering (CODC) algorithm iteratively selects target domain clusters T_k and source training sets S_k, which are allowed to differ. The paper reports consistent DomainAUC improvements on Amazon and AliCCP, offline industrial results on an app marketplace, and a 4.9% online eCPM lift over 14 days across 64 domains.

Significance. If the central claims hold, CDC makes a useful contribution to multi-domain recommendation at scale: it explicitly decouples source training domains from target inference domains, it is model-agnostic and can wrap existing MDR models, and the public-dataset evaluations include significance tests over five runs with released code. The paper also reports a large online A/B test, which is rare and valuable. However, the signature causal-integration mechanism is currently under-tested: no experiment isolates the functional form of the lambda coefficient in Eq. (10), the whole-method ablation does not distinguish the causal distance metric from the integration rule, and Algorithm 2 leaves key update conditions unspecified. The significance is therefore real but conditional on tightening the evidence for the causal component and on clarifying the algorithmic details.

major comments (4)
  1. [Section 4.3, Eq. (10)] The domain interaction coefficient lambda_{d_v|S} is asserted as min(1, sum_{d_u in S} M_C[u,v] / (2|S| c_S)) without a derivation from the causal framework of Markham et al. [21] or from the lookahead estimates. The hypotheses in Section 4.3 justify only the sign and monotonicity of the dependence on cohesion, not the specific ratio, the min(1, .) cap, or the reuse of the same coefficient in Eq. (6), Eq. (14), and Algorithm 2. The ablation labeled 'w/o Causal' in Figure 3a replaces the kernel in Eq. (8) with cosine similarity, which tests the distance metric, not the lambda integration logic; no ablation fixes lambda to a constant or otherwise misspecifies Eq. (10). As written, the claim that causal discovery is what makes the dual-view integration work is underdetermined by the presented experiments.
  2. [Algorithm 2, lines 2-4 and line 9] The outer loop condition 'if condition to update domain matrices is met' is never defined, and the convergence criterion for the 'while CDC has not converged' loop is not specified. The frequency of lookahead updates controls the accuracy and cost of M_I, M_H, and M_T, and the paper's feedback mechanism depends on this schedule. In addition, line 9 says T_k^t is initialized from domains in T_k^{t-1} where lambda is 'the smallest', but does not state how many such domains are retained. These unspecified steps prevent exact reproduction of CODC, which is a central contribution.
  3. [Section 3.1, Eqs. (1)-(2)] Equation (1) defines the parameters Theta_{S_k} of the model trained on S_k as minimizing the loss summed over d in T_k, which is internally inconsistent with the stated source-set/target-set separation. The training objective should presumably be over S_k, with T_k used for evaluation. As written, the formal problem definition conflates the two sets and undermines the precise statement of the paper's main innovation.
  4. [Table 4 and Section 5.6] The industrial offline results in Table 4 are reported as deltas in per-mille relative to the DCN baseline, but no error bars, number of runs, or significance tests are given. Given that several entries are small relative differences and at least one row (CDC-S_k) appears to contain a formatting error ('+2.85‰-+2.65‰2.65‰'), the strength of the industrial offline evidence is not commensurate with the claims. Please state how many runs were performed and provide confidence intervals or a statistical test, or clearly mark these as single-run measurements.
minor comments (5)
  1. [Table 4, CDC-S_k row] The entry '+2.85‰-+2.65‰2.65‰' appears to be a typographical or formatting error; please correct it so the reported values are unambiguous.
  2. [Section 4.4.1, text near Eq. (14)] The phrase 'provides a reliable and robust reliable starting point' contains a duplicated word and should be revised.
  3. [Figure 3, subfigure (a)] The text in Section 5.7 says the leftmost subplot reports DomainAUC, but the y-axis values shown (0.654-0.675) correspond more closely to the Minor10AUC values in Table 2; please clarify which metric is plotted and label the axis accordingly.
  4. [Section 4.3.1, Eq. (7)] The construction of the random domain sets R_r^t used as treatments is not described (e.g., how many domains are sampled per action and how the sample proportions are chosen), although these sets determine the treatment effect matrix M_T; please add the missing details.
  5. [Section 5.4 and Table 5] The complexity analysis in Table 5 reports 'OOM' for the Multi row on MDR-229M but does not state the hardware or memory budget; a brief note would make the comparison more interpretable.

Circularity Check

0 steps flagged · score 2.0 of 10

No load-bearing circularity: CDC's transfer-gain construction is measured and externally evaluated; only non-load-bearing background self-citations appear.

full rationale

I walked the derivation chain from the two affinity matrices (Eqs. 3-4) through the treatment-effect matrix M_T (Eq. 7), the causal-distance matrix M_C (Eq. 8), the cohesion and interaction coefficient (Eqs. 9-10), and the transfer gain J (Eq. 6). At no point is a target performance quantity such as DomainAUC or eCPM used to fit the components of J. Eq. (10) defines lambda as min(1, sum_{u in S} M_C[u,v] / (2|S| c_S)), with c_S and M_C computed from lookahead loss measurements; Eq. (6) is a fixed weighted combination of M_I and M_H. The held-out AUC results in Tables 2-3 and the 14-day online A/B test are independent of these constructions, so the reported gains cannot reduce by construction to the affinities that produced them. The feedback between S_k and M_H expressed in Eq. (5) is an iterative co-optimization, not a logical circularity: the final clustering is evaluated externally on unseen data. The principal weakness is that Eq. (10) is an asserted heuristic rather than a derivation from the cited causal framework, and the w/o Causal ablation replaces the kernel in Eq. (8) rather than isolating the functional form of lambda; this is underdetermination of a component, not circularity. The only author-overlap citations are [18] and [40], used for background definitions of causal-discovery assumptions and negative transfer; neither is load-bearing, because the causal-distance component relies on the external kernel of Markham et al. [21] and the headline results are externally benchmarked. No fitted parameter is renamed as a prediction, and no equation is equivalent to its own input by construction. The score of 2 reflects only the presence of minor, non-load-bearing self-citations.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The ledger shows that the method depends on several heuristic modeling choices (cohesion, lambda, source-set initialization) rather than a single fitted constant. The causal framing is borrowed from an external kernel, and the transfer measurements themselves are treated as ground truth without an external benchmark.

free parameters (6)
  • alpha (smoothing factor) = 0.3
    Smooths historical affinity matrices in Eqs. (3) and (4); tuned on Amazon validation.
  • rho0 (initial affiliation weight) = 0.01
    Initial weight for P(d_u|k) in Eq. (15); tuned on Amazon validation.
  • beta (decay rate) = 0.95
    Decay rate for rho_t in Eq. (15); tuned on Amazon validation.
  • K (number of clusters) = 3 (Amazon)
    Searched in [2,3,4]; optimal value depends on dataset.
  • N (initial source set size) = 2
    Initial size of S_k,0 in Eq. (12); set by hand.
  • Learning rates and batch sizes = lr in [5e-4,1e-3,2e-3,3e-3], batch in [1024,2048,4096]
    Grid-searched for Adam on public datasets; not reported whether identical ranges were used for all baselines.
assumptions (4)
  • domain assumption One-step lookahead loss ratios reflect post-training transfer effectiveness.
    All affinity matrices (Eqs. 3, 4, 7) and the CODC algorithm rely on this to define transfer; the paper motivates it with Challenge 1 but does not prove it.
  • standard math The dependence contribution kernel kappa is an appropriate distance in the transfer-effect space.
    Imported from Markham et al. (2022); used in Eq. (8) to compute causal distance from treatment-effect vectors.
  • ad hoc to paper The cohesion coefficient formula lambda_{d_v|S} correctly weights isolated vs hybrid affinities.
    Eq. (10) is introduced with two informal hypotheses in Section 4.3; no derivation from the causal framework is given.
  • domain assumption Domains have no true causal relations, yet treatment-effect vectors are sufficient statistics for domain interdependence.
    Stated explicitly in Section 4.3.1; the entire 'causal distance' construction depends on this, and it is not independently verified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CDC: Causal Domain Clustering for Multi-Domain Recommendation." pith.science (2026). https://pith.science/paper/DQ5S72M6

@misc{pith2026250706877,
  author       = {Pith},
  title        = {Pith review of: CDC: Causal Domain Clustering for Multi-Domain Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DQ5S72M6}},
  note         = {Machine review of arXiv:2507.06877}
}
read the original abstract

Multi-domain recommendation leverages domain-general knowledge to improve recommendations across several domains. However, as platforms expand to dozens or hundreds of scenarios, training all domains in a unified model leads to performance degradation due to significant inter-domain differences. Existing domain grouping methods, based on business logic or data similarities, often fail to capture the true transfer relationships required for optimal grouping. To effectively cluster domains, we propose Causal Domain Clustering (CDC). CDC models domain transfer patterns within a large number of domains using two distinct effects: the Isolated Domain Affinity Matrix for modeling non-interactive domain transfers, and the Hybrid Domain Affinity Matrix for considering dynamic domain synergy or interference under joint training. To integrate these two transfer effects, we introduce causal discovery to calculate a cohesion-based coefficient that adaptively balances their contributions. A Co-Optimized Dynamic Clustering algorithm iteratively optimizes target domain clustering and source domain selection for training. CDC significantly enhances performance across over 50 domains on public datasets and in industrial settings, achieving a 4.9% increase in online eCPM. Code is available at https://github.com/Chrissie-Law/Causal-Domain-Clustering-for-Multi-Domain-Recommendation

Figures

Figures reproduced from arXiv: 2507.06877 by the authors.

Figure 1
Figure 1. Compared with existing MDR and domain group [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Causal Domain Clustering (CDC) framework. (1) We learn two inter-domain affinity matrices, [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Ablation study (subfigure (a)) and hyper-parameter [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Case study: Compared to manual grouping based on [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 26 canonical work pages

  1. [21]

    Alex Markham, Richeek Das, and Moritz Grosse-Wentrup. 2022. A distance covariance-based kernel for nonlinear causal clustering in heterogeneous popu- lations. InConference on Causal Learning and Reasoning. PMLR, 542–558

    work page 2022

  2. [1]

    Guangji Bai and Liang Zhao. 2022. Saliency-Regularized Deep Multi-Task Learn- ing. InProceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining(Washington DC, USA)(KDD ’22). Association for Computing Machinery, New York, NY, USA, 15–25. doi:10.1145/3534678.3539442

    work page arXiv 2022

  3. [2]

    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

  4. [3]

    Inderjit S Dhillon, Yuqiang Guan, and Brian Kulis. 2004. Kernel k-means: spec- tral clustering and normalized cuts. InProceedings of the tenth ACM SIGKDD international conference on Knowledge discovery and data mining. 551–556

    work page 2004

  5. [4]

    Inderjit S Dhillon, Edward M Marcotte, and Usman Roshan. 2003. Diametrical clustering for identifying anti-correlated gene clusters.Bioinformatics19, 13 (2003), 1612–1619

    work page 2003

  6. [5]

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

    2006

  7. [6]

    Chris Fifty, Ehsan Amid, Zhe Zhao, Tianhe Yu, Rohan Anil, and Chelsea Finn

    work page

  8. [7]

    Huifeng Guo, Ruiming Tang, Yunming Ye, Zhenguo Li, and Xiuqiang He. 2017. DeepFM: a factorization-machine based neural network for CTR prediction. In Proceedings of the 26th International Joint Conference on Artificial Intelligence. 1725–1731

    2017

Show all 44 references
  1. [8]

    David Ha, Andrew Dai, and Quoc V Le. 2016. Hypernetworks.arXiv preprint arXiv:1609.09106(2016)

    2016 arXiv

  2. [9]

    Vishwanath R Iyer, Michael B Eisen, Douglas T Ross, Greg Schuler, Troy Moore, Jeffrey CF Lee, Jeffrey M Trent, Louis M Staudt, James Hudson Jr, Mark S Boguski, et al. 1999. The transcriptional program in the response of human fibroblasts to serum.science283, 5398 (1999), 83–87

    1999

  3. [10]

    Pengyue Jia, Yichao Wang, Shanru Lin, Xiaopeng Li, Xiangyu Zhao, Huifeng Guo, and Ruiming Tang. 2024. D3: A Methodological Exploration of Domain Division, Modeling, and Balance in Multi-Domain Recommendations. InProceedings of the AAAI Conference on Artificial Intelligence, Vo...

    2024

  4. [11]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. 2015. Adam: A Method for Stochastic Opti- mization. In3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, Yoshua Bengio and Yann LeCun (Eds.). http://arxiv.org/ab...

    2015 arXiv

  5. [12]

    Jinyun Li, Huiwen Zheng, Yuanlin Liu, Minfang Lu, Lixia Wu, and Haoyuan Hu. 2023. ADL: Adaptive Distribution Learning Framework for Multi-Scenario CTR Prediction. InProceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval....

    2023

  6. [13]

    Qingyun Liu, Zhe Zhao, Liang Liu, Zhen Zhang, Junjie Shan, Yuening Li, Shuchao Bi, Lichan Hong, and Ed H Chi. 2023. Multitask Ranking System for Immersive Feed and No More Clicks: A Case Study of Short-Form Video Recommendation. InProceedings of the 32nd ACM International Conf...

    2023

  7. [14]

    Qi Liu, Zhilong Zhou, Gangwei Jiang, Tiezheng Ge, and Defu Lian. 2023. Deep task-specific bottom representation network for multi-task recommendation. InProceedings of the 32nd ACM International Conference on Information and Knowledge Management. 1637–1646

    2023

  8. [15]

    Shikun Liu, Stephen James, Andrew J Davison, and Edward Johns. 2022. Auto- Lambda: Disentangling Dynamic Task Relationships.Transactions on Machine Learning Research(2022)

    2022

  9. [16]

    Shengchao Liu, Yingyu Liang, and Anthony Gitter. 2019. Loss-balanced task weighting to reduce negative transfer in multi-task learning. InProceedings of the AAAI conference on artificial intelligence, Vol. 33. 9977–9978

    2019

  10. [17]

    Zhi-Ping Liu. 2015. Reverse engineering of genome-wide gene regulatory net- works from gene expression data.Current genomics16, 1 (2015), 3–22

    2015

  11. [18]

    Huishi Luo, Fuzhen Zhuang, Ruobing Xie, Hengshu Zhu, Deqing Wang, Zhulin An, and Yongjun Xu. 2024. A survey on causal inference for recommendation. The Innovation5, 2 (2024), 100590. doi:10.1016/j.xinn.2024.100590

    2024

  12. [19]

    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

  13. [20]

    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. InThe 41st International ACM SIGIR Conference on Research & Development in Information Retriev...

    2018

  14. [22]

    Jianmo Ni, Jiacheng Li, and Julian McAuley. 2019. Justifying recommendations using distantly-labeled reviews and fine-grained aspects. InProceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural...

    2019

  15. [23]

    2009.Causality(2 ed.)

    Judea Pearl. 2009.Causality(2 ed.). Cambridge University Press. doi:10.1017/ CBO9780511803161

    2009

  16. [24]

    Dripta S Raychaudhuri, Yumin Suh, Samuel Schulter, Xiang Yu, Masoud Faraki, Amit K Roy-Chowdhury, and Manmohan Chandraker. 2022. Controllable dy- namic multi-task architectures. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 10955–10964

    2022

  17. [25]

    Qijie Shen, Wanjie Tao, Jing Zhang, Hong Wen, Zulong Chen, and Quan Lu. 2021. SAR-Net: A scenario-aware ranking network for personalized fair recommenda- tion in hundreds of travel scenarios. InProceedings of the 30th ACM International Conference on Information & Knowledge Man...

    2021

  18. [26]

    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 prediction. In Proceedings of the 30th ACM Internationa...

    2021

  19. [27]

    Ammar Sherif, Abubakar Abid, Mustafa Elattar, and Mohamed ElHelw. 2024. STG- MTL: scalable task grouping for multi-task learning using data maps.Machine Learning: Science and Technology5, 2 (2024), 025068

    2024

  20. [28]

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

    2019

  21. [29]

    Xiaozhuang Song, Shun Zheng, Wei Cao, James Yu, and Jiang Bian. 2022. Effi- cient and effective multi-task grouping via meta learning on task combinations. Advances in Neural Information Processing Systems35 (2022), 37647–37659

    2022

  22. [30]

    Trevor Standley, Amir Zamir, Dawn Chen, Leonidas Guibas, Jitendra Malik, and Silvio Savarese. 2020. Which tasks should be learned together in multi-task learning?. InInternational conference on machine learning. PMLR, 9120–9132

    2020

  23. [31]

    Swabha Swayamdipta, Roy Schwartz, Nicholas Lourie, Yizhong Wang, Hannaneh Hajishirzi, Noah A Smith, and Yejin Choi. 2020. Dataset cartography: Mapping and diagnosing datasets with training dynamics.arXiv preprint arXiv:2009.10795 (2020)

    2020 arXiv

  24. [32]

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

    2020

  25. [33]

    Chenguang Wang, Xuanhao Pan, and Tianshu Yu. 2024. Towards Principled Task Grouping for Multi-Task Learning.arXiv preprint arXiv:2402.15328(2024)

    2024 arXiv

  26. [34]

    Ruoxi Wang, Bin Fu, Gang Fu, and Mingliang Wang. 2017. Deep & cross network for ad click predictions. InProceedings of the ADKDD’17. 1–7

    2017

  27. [35]

    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. 1785–1797

    2021

  28. [36]

    Yichao Wang, Huifeng Guo, Bo Chen, Weiwen Liu, Zhirong Liu, Qi Zhang, Zhicheng He, Hongkun Zheng, Weiwei Yao, Muyu Zhang, et al. 2022. Causalint: Causal inspired intervention for multi-scenario recommendation. InProceedings of the 28th ACM SIGKDD Conference on Knowledge Discov...

    2022

  29. [37]

    Xuanhua Yang, Xiaoyu Peng, Penghui Wei, Shaoguo Liu, Liang Wang, and Bo Zheng. 2022. Adasparse: Learning adaptively sparse structures for multi-domain click-through rate prediction. InProceedings of the 31st ACM International Con- ference on Information & Knowledge Management....

    2022

  30. [38]

    Yuanliang Zhang, Xiaofeng Wang, Jinxin Hu, Ke Gao, Chenyi Lei, and Fei Fang

  31. [39]

    Jie Zhou, Xianshuai Cao, Wenhao Li, Lin Bo, Kun Zhang, Chuan Luo, and Qian Yu

  32. [40]

    Fuzhen Zhuang, Zhiyuan Qi, Keyu Duan, Dongbo Xi, Yongchun Zhu, Hengshu Zhu, Hui Xiong, and Qing He. 2020. A comprehensive survey on transfer learning. Proc. IEEE109, 1 (2020), 43–76

    2020

  33. [41]

    Xinyu Zou, Zhi Hu, Yiming Zhao, Xuchu Ding, Zhongyi Liu, Chenliang Li, and Aixin Sun. 2022. Automatic expert selection for multi-scenario and multi-task search. InProceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1535–1544

    2022

  34. [2021]

    InAd- vances in Neural Information Processing Systems, M

    Efficiently Identifying Task Groupings for Multi-Task Learning. InAd- vances in Neural Information Processing Systems, M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan (Eds.), Vol. 34. Curran Associates, Inc., 27503–27516. https://proceedings.neurips....

    2021

  35. [2022]

    InProceedings of the 31st ACM International Conference on Information & Knowledge Management

    Scenario-adaptive and self-supervised model for multi-scenario personal- ized recommendation. InProceedings of the 31st ACM International Conference on Information & Knowledge Management. 3674–3683

  36. [2023]

    In2023 IEEE 39th International Conference on Data Engineering (ICDE)

    Hinet: Novel multi-scenario & multi-task learning with hierarchical infor- mation extraction. In2023 IEEE 39th International Conference on Data Engineering (ICDE). IEEE, 2969–2975

Pith tools

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