Pith. sign in

REVIEW 5 major objections 7 minor 45 references

MERIT: A Merchant Incentive Ranking Model for Hotel Search & Ranking

T0 review · 5 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read MERIT adds a monotonic merchant-quality tower to hotel ranking, lifting displayed-hotel MCI by 3.02% online while holding consumer metrics.

desk verdict A deployed industry paper with a real idea and a real circularity problem: MCI is both the model's input and its training/evaluation target, so the impressive NDCG and online MCI lifts are partly self-measured. read the letter →

arxiv 2506.08442 v1 pith:ILNGNGLK submitted 2025-06-10 cs.IR

Shigang Quan , Hailong Tan , Shui Liu , Zhenzhe zheng , Ruihao Zhu , Liangyue Li , Quan Lu , Fan Wu This is my paper · ORCID
classification cs.IR
keywords hotelsearchandrankingmerchantincentiveCompetitivenessIndexmonotonicneuralnetworkmulti-objectivepairwiselossmulti-tasklearningonlinetravelplatform
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

This paper tries to establish that a hotel search-and-ranking system can deliberately favor higher-quality merchants without hurting consumers, by making merchant quality an explicit, monotonic input to the ranking score. The authors define a Merchant Competitiveness Index (MCI) from merchant operational and service indicators, feed it through a 'Merchant Tower' whose weights are constrained positive, so better MCI factors can only raise a hotel's rank. A Multi-objective Stratified Pairwise Loss subordinates MCI ranking to the click/order objective, so quality is rewarded only when it does not conflict with consumer conversion. The claim is supported by an offline NDCG@5 for MCI ranking of 0.8718 versus 0.8468 for a DNN baseline, and by an online A/B test showing MCI +3.02%, CVR +1.50%, and sales volume +0.97%. If right, this gives travel platforms a concrete mechanism for closing an incentive loop that tilts impressions to good merchants and motivates them to improve service.

What carries the argument

The two load-bearing objects are the Merchant Tower and the Multi-objective Stratified Pairwise Loss (MSPL). The Merchant Tower is a positive-weight feed-forward network $\phi^B_\theta(\cdot)$ that takes the MCI feature vector $X_s$ along with the CTR/CVR representations, with weights constrained so that increasing any MCI factor cannot lower the ranking score; this gives merchants a legible, monotonic quality-to-position rule. MSPL is the loss $$$L^{{MCI|CTRCVR}}$_{pairwise} = \sum_{i,j} I(y_i \ge y_j) \ell(\hat{s}_i - \hat{s}_j, I(z_i > z_j)),$$ which masks MCI pairwise updates whenever the primary click/order label disagrees, resolving the gradient conflict between short-term revenue and long-term quality. The Merchant Incentive Module combines these with a Deep&Cross Network and entire-space modeling $pCTCVR = pCTR \times pCVR$, letting the model carry merchant quality as a first-class signal while keeping the consumer objective primary.

What would settle it

Run the offline evaluation with the MCI factors removed from the input feature set, so $x_s$ no longer contains the indicators that define the label $z$. If the NDCG@5 gain over DNN shrinks toward zero, the reported ranking improvement is mostly the model predicting the label from its own causes, not evidence that the monotonic Merchant Tower and MSPL create a genuine quality incentive.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that a hotel ranking model can carry both consumer and merchant objectives in a single differentiable architecture: a standard CTR/CVR prediction pipeline with Deep&Cross features and entire-space modeling of $pCTCVR = pCTR \times pCVR$ is augmented with a monotonic Merchant Tower that maps the MCI feature vector to a score, and a stratified pairwise loss trains the MCI ranking only on sample pairs where the primary click/order label is not contradicted. The authors report that this design lifts offline MCI-ranking NDCG@5 from 0.8468 to 0.8718 over a DNN baseline, and that in a two-week online A/B test the average MCI of displayed hotels rose 3.02% while CVR rose 1.50% and sales volume 0.97%. They interpret these results as evidence that the Matthew Effect in consumer feedback can be counteracted, that merchants receive a clear monotonic relation between quality factors and ranking performance, and that short-term revenue need not be sacrificed to improve long-term quality signals.

Load-bearing premise

The approach assumes that the Merchant Competitiveness Index is a trustworthy ground-truth measure of hotel quality, even though the same indicator factors are fed into the model as input features and also used as the training and evaluation label.

Editorial extensions

If this is right

  • Merchants get an explicit, monotonic rule: improving any MCI factor, such as lower order-refusal rate or better picture quality, can only improve their ranking score, turning quality signals into actionable investment targets.
  • Platforms can tune the loss weights $\lambda_1$ and $\lambda_2$ to choose a tolerable conversion loss; the paper treats a 0.005 drop in CTCVR AUC as acceptable while gaining 0.0250 in MCI-ranking NDCG@5.
  • In the online A/B test the model reallocated impressions toward higher-MCI hotels, lifting UCVR for top-quality hotels (for example +4.93% in Chengdu at MCI level 5.0), evidence that the incentive loop can operate at scale.
  • Multi-task baselines that ignore merchant quality do not improve MCI ranking; the gains come from the monotonic tower plus the stratified MCI pairwise loss, not from multi-task sharing alone.

Reading between the lines

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

  • If MCI is partly manipulable, making it monotonic could incentivize merchants to game the index components rather than genuinely improve service; a field experiment that discloses the MCI mapping and then tracks actual service investments would test this.
  • The stratified pairwise loss is a general template: any two-tier objective with engagement as the primary label and a quality, safety, or fairness score as the secondary label could adopt the same masking to avoid gradient conflict.
  • The online MCI lift may include a position-bias component: high-MCI hotels placed higher get more clicks, which raises their historical conversion rates and hence their future MCI, potentially recreating the Matthew Effect along a new axis. A longer-horizon A/B test with counterfactual position assignments would separate this feedback from the direct ranking effect.
  • The offline NDCG gain should be re-examined with the MCI-defining factors held out of the input features, because the same Table 1 indicators are used as both inputs and the training label.
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 / 7 minor

Summary. This paper proposes MERIT, a hotel search ranking model that explicitly incorporates hotel merchant quality through a Merchant Competitiveness Index (MCI). The model consists of a Deep&Cross network with CTR/CVR towers plus a Merchant Tower that is constrained to be monotone in the MCI features, and a Multi-objective Stratified Pairwise Loss (MSPL) that balances consumer feedback (click/order) with MCI ranking. Offline experiments on a large Fliggy dataset report NDCG improvements for MCI ranking (NDCG@5 0.8718 vs 0.8468 for DNN) while consumer AUC metrics are roughly flat or slightly worse. An online A/B test reports +3.02% mean MCI score, +1.50% CVR, and +0.97% sales volume. The authors claim the method has been deployed on Fliggy.

Significance. The problem is practically important: aligning merchant incentives with platform ranking can improve the long-term health of online travel platforms. The technical components, including a monotone Merchant Tower and a stratified pairwise loss, are reasonable and potentially reusable. However, the evaluation as presented does not convincingly establish the central claim. The MCI is used both as an input feature and as the training/evaluation label, which makes the offline NDCG gains partly a self-consistency result, and the online MCI lift is the metric being optimized. Hyperparameters are selected on the test set, further biasing the reported results. The paper's strengths are its large-scale real-world dataset, the reported deployment, and the clear description of the model architecture; these are undermined by the evaluation gaps.

major comments (5)
  1. [Section 3.2, Section 4.2, Eq. (9)/(11)] The central offline evidence is weakened by the circular use of MCI. Section 3.2 states that the MCI score is used for both the input feature x_MCI and the ranking label z, and Section 4.2 feeds the Table 1 indicators X_s into the Merchant Tower through a positive-weight network phi_B, guaranteeing monotonicity in X_s. Since z is presumably computed from the same indicators, the model can recover z from X_s almost by construction. The NDCG gain of MERIT+MSPL over DNN in Table 3 therefore largely reflects that MERIT is explicitly trained on the MCI ranking objective while the baselines are not, rather than establishing a merchant-aware ranking mechanism. Please address this by, for example, training all baselines with the same MCI pairwise objective, holding out the MCI computation from a validation set, or evaluating against an external merchant-quality label.
  2. [Section 5.3] The hyper-parameters lambda1 and lambda2 are selected on the test set. The text describes choosing the point with the largest NDCG@20 among those satisfying a CTCVR AUC lower bound, where both metrics are computed on the test set of the offline dataset. This test-set-based model selection inflates the reported NDCG numbers and invalidates them as unbiased estimates. The selection should be done on a validation split, with the test set used only for final evaluation.
  3. [Table 2 vs. Section 5.4] The offline consumer metrics show MERIT+MSPL slightly underperforming DNN on CTR AUC (-0.0002), CVR AUC (-0.0046), and CTCVR AUC (-0.0023), yet the online A/B test reports +1.50% CVR with no confidence interval, p-value, or experimental scale. Given the offline decrease, the online positive result requires statistical detail (e.g., number of users, duration, significance test, or day-level consistency) to be credible.
  4. [Section 5.4] The headline online improvement of +3.02% is the mean MCI score of displayed hotels, which is exactly the quantity MERIT is trained to maximize. This is a self-consistency result, not evidence of improved merchant quality or of the claimed incentive loop. The paper should report independent outcome metrics, such as review scores, repeat-booking rates, or consumer complaint rates, or explicitly acknowledge that the online test does not measure merchant quality independently.
  5. [Section 3.2, Table 1] The exact formula for the MCI score is never disclosed. The paper lists indicators in Table 1 but does not specify how they are aggregated into a single score z. Since z is the training label and the evaluation target for NDCG, the reader cannot assess the validity of the label or the degree of circularity between X_s and z. Provide the aggregation formula or, if it is proprietary, at least describe its functional form and the source of the weights.
minor comments (7)
  1. [Abstract] The phrase 'MERchant InceTive' contains a typo and should read 'MERchant Incentive'.
  2. [Section 4.1, Eq. (2)] The variable r is defined but never used, and the MCI embedding E_s is omitted from the concatenated vector E even though X_s is later fed into the Merchant Tower; please clarify the notation.
  3. [Section 5.1] It is unclear whether the DNN and multi-task baselines also receive X_s (the MCI features) as inputs and whether they are trained with any MCI objective; state this explicitly for a fair comparison.
  4. [Figure 4] The axis labels are garbled (e.g., '1 = 0.1' should be 'λ1 = 0.1'), and the dashed line and light green area are difficult to distinguish in grayscale.
  5. [Section 5.4] The experiment window 'within two weeks of July 2022' is vague; specify the exact start and end dates.
  6. [Conclusion] The statement that MERIT outperforms state-of-the-art benchmarks should be qualified to MCI ranking, since the consumer metrics are not consistently better.
  7. [References] Reference [30] has an incomplete citation ('In SSRN 3670264'); provide the full title and year.

Circularity Check

3 steps flagged · score 7.0 of 10

Central merchant-side evidence is circular: MCI is used as both input feature and ranking label, so offline NDCG@5 and online MCI +3.02% measure fitting the MCI score from its own component factors.

  1. self definitional [Section 3.2 (Merchant Quality Rating Score)]
    "In this paper, the MCI score will be used for both input feature 𝑥𝑀𝐶𝐼 in Section 3.3 and ranking label 𝑧 in Section 4.1."

    The paper defines MCI and then explicitly declares it to be both a model input and the model's ranking label. Any model that receives the Table 1 MCI factors as inputs and is trained to predict z can recover a large part of z by learning the undisclosed MCI formula from its own inputs. Consequently, Table 3's NDCG@5 on the MCI label is a self-consistency measure, not independent evidence that MERIT discovers or incentivizes merchant quality.

  2. fitted input called prediction [Section 4.1 and Equation (9)]
    "The features of MCI in Table 1 will be represented as 𝑥𝑠."

    The Merchant Tower receives X_s, the same indicators used to construct the MCI label, and Equation (9) trains a pairwise ranking loss on z with the stratified term I(y_i >= y_j). A monotone tower in X_s can therefore reproduce a monotone transform of z from its own causes. The 0.0250 NDCG@5 improvement over DNN in Table 3 is partly an objective-mismatch artifact: DNN and the multi-task baselines are never trained with an MCI-ranking objective, while MERIT+MSPL optimizes exactly that target. The claimed 'clear relation between hotel quality and performance' is partly built in rather than learned.

1 more flagged steps
  1. fitted input called prediction [Section 5.4 (Online A/B Test)]
    "The mean MCI score of displayed hotels has risen by 3.02%, indicating that displayed merchant quality has been improved."

    The online headline metric is the mean MCI score of displayed hotels, which is precisely the quantity that L_MCI|CTRCVR in Equation (9) trains MERIT to increase. Measuring the A/B test on the same self-defined score that the model optimizes does not validate merchant quality independently. Without an external quality label or a long-run consumer outcome tied to the incentive loop, the +3.02% MCI lift is an expected consequence of optimizing for z, not evidence of a genuine merchant-incentive mechanism.

full rationale

The central circularity is self-definitional rather than citation-based. Section 3.2 states without ambiguity that the MCI score is used as both the input feature x_MCI and the ranking label z, and Section 4.1 confirms that the model's X_s features are the Table 1 MCI factors. Equation (9) then trains a pairwise ranking loss on z, and the Merchant Tower is constrained to be monotone in X_s. As a result, the offline NDCG@5/wNDCG gains in Table 3 largely measure the model's ability to fit the MCI score from its own component indicators, and the online MCI lift in Table 5 measures an increase in the very objective being optimized. The comparison is also asymmetric: DNN, Shared Bottom, MMoE, and CGC are not given an MCI-ranking loss, so part of the reported gain reflects objective mismatch rather than a superior merchant-aware ranking mechanism. The paper does not disclose the MCI formula, and it does not include a limitation section acknowledging this input-label coupling. I do not see a load-bearing self-citation chain or imported uniqueness theorem; the circularity is internal to the MCI construction. There is some independent content: the consumer-side CTR/CVR metrics and the online CVR (+1.50%) and sales volume (+0.97%) improvements are external to the MCI label and are not forced by the MCI objective. For that reason the paper is not wholly circular, but its central merchant-side claim does reduce by construction to optimizing a self-defined score, warranting a score of 7 rather than 8 or 10.

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

MCI is the central construct and it appears on both sides of the learning problem. The business-defined index is never validated against external outcomes, its aggregation weights are not disclosed, and the model assumes monotonicity of quality-to-score. Hyperparameters lambda1 and lambda2 are tuned on the test period. These ledger entries capture the main costs a reader must accept.

free parameters (4)
  • lambda_1 = 1.0
    Pairwise CTR/CVR loss weight tuned by grid search on offline test metrics in Section 5.3.
  • lambda_2 = 0.1
    Pairwise MCI loss weight tuned by grid search, maximizing NDCG@20 subject to a CTCVR AUC threshold.
  • acceptable_CTCVR_AUC_decline = 0.005
    Business tolerance threshold chosen from online experience and used to select lambda values in Section 5.3.
  • MCI_aggregation_weights = not disclosed
    The paper lists MCI factors in Table 1 but does not specify how Operational Capability, Room Inventory, Quality of Service, and Basic Information Score are weighted into the MCI score.
assumptions (4)
  • domain assumption MCI is a valid, complete measure of merchant quality.
    Section 3.2 Table 1 defines MCI from GMV, inventory ratios, refusal rates, and so on. No independent validation links MCI to long-term consumer value or merchant contribution.
  • ad hoc to paper Ranking score should be monotonically nondecreasing in MCI features.
    The Merchant Tower uses a positive-weight network over X_s (Equations 4 and 5) to enforce monotonicity. This assumes higher quality always warrants better ranking for every query and consumer.
  • ad hoc to paper MCI can be used both as model input and as training/evaluation label without invalidating results.
    Section 3.2 states the MCI score is used as input feature x_MCI and label z. If x_s includes the same factors that define z, NDCG gains partly measure self-consistency.
  • domain assumption Entire-space modeling assumptions from ESMM apply to hotel search and ranking.
    Section 4.2 follows ESMM to avoid sample selection bias and data sparsity, assuming click and conversion are sequential and observable for all impressions.
invented entities (1)
  • Merchant Competitiveness Index (MCI)
    purpose: A composite score representing hotel merchant quality, used as both an input feature and a ranking label.
    Introduced in Section 3.2 using four weighted indicator groups. No external benchmark validates it, and the online MCI lift is measured against the same construct the model optimizes.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MERIT: A Merchant Incentive Ranking Model for Hotel Search & Ranking." pith.science (2026). https://pith.science/paper/ILNGNGLK

@misc{pith2026250608442,
  author       = {Pith},
  title        = {Pith review of: MERIT: A Merchant Incentive Ranking Model for Hotel Search & Ranking},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ILNGNGLK}},
  note         = {Machine review of arXiv:2506.08442}
}
read the original abstract

Online Travel Platforms (OTPs) have been working on improving their hotel Search & Ranking (S&R) systems that facilitate efficient matching between consumers and hotels. Existing OTPs focus almost exclusively on improving platform revenue. In this work, we take a first step in incorporating hotel merchants' objectives into the design of hotel S&R systems to achieve an incentive loop: the OTP tilts impressions and better-ranked positions to merchants with high quality, and in return, the merchants provide better service to consumers. Three critical design challenges need to be resolved to achieve this incentive loop: Matthew Effect in the consumer feedback-loop, unclear relation between hotel quality and performance, and conflicts between short-term and long-term revenue. To address these challenges, we propose MERIT, a MERchant IncenTive ranking model, which can simultaneously take the interests of merchants and consumers into account. We define a new Merchant Competitiveness Index (MCI) to represent hotel merchant quality and propose a new Merchant Tower to model the relation between MCI and ranking scores. Also, we design a monotonic structure for Merchant Tower to provide a clear relation between hotel quality and performance. Finally, we propose a Multi-objective Stratified Pairwise Loss, which can mitigate the conflicts between OTP's short-term and long-term revenue. The offline experiment results indicate that MERIT outperforms these methods in optimizing the demands of consumers and merchants. Furthermore, we conduct an online A/B test and obtain an improvement of 3.02% for the MCI score.

Figures

Figures reproduced from arXiv: 2506.08442 by the authors.

Figure 1
Figure 1. The positive incentive loop for hotel merchants on [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The overview framework of hotel S&R system. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The overview architecture of MERIT, which consists of the Feature Representation Layer, Concatenate Layer, Merchant [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: The NDCG@20 of MCI ranking and CTCVR AUC for different hyper-parameters [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: The online performance of MERIT model for ho [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

45 extracted references · 23 canonical work pages

  1. [1]

    Norman P Archer and Shouhong Wang. 1993. Application of the back propa- gation neural network algorithm with monotonicity constraints for two-group classification problems. Decision Sciences 24, 1 (1993), 60–75

    work page 1993

  2. [2]

    Oren Barkan and Noam Koenigstein. 2016. Item2vec: neural item embedding for collaborative filtering. In 2016 IEEE 26th International Workshop on Machine Learning for Signal Processing (MLSP) . IEEE, 1–6

    work page 2016

  3. [3]

    Rajat Bhattacharjee and Ashish Goel. 2006. Incentive based ranking mechanisms. In First Workshop on the Economics of Networked Systems (Netecon’06) . 62–68

    work page 2006

  4. [4]

    Rajat Bhattacharjee, Ashish Goel, and Konstantinos Kollias. 2009. An incentive- based architecture for social recommendations. In Proceedings of the third ACM conference on Recommender systems . 229–232. CIKM ’23, October 21–25, 2023, Birmingham, United Kingdom. Quan, et al

    work page 2009

  5. [5]

    Canini, A

    K. Canini, A. Cotter, M. R. Gupta, M. Milani Fard, and J. Pfeifer. 2016. Fast and Flexible Monotonic Functions with Ensembles of Lattices. In Proceedings of the 30th International Conference on Neural Information Processing Systems (NIPS’16) . 2927–2935

    work page 2016

  6. [6]

    Rich Caruana. 1997. Multitask learning. Machine learning 28, 1 (1997), 41–75

    work page 1997

  7. [7]

    Jiawei Chen, Hande Dong, Xiang Wang, Fuli Feng, Meng Wang, and Xiangnan He. 2020. Bias and debias in recommender system: A survey and future directions. arXiv preprint arXiv:2010.03240 (2020)

    arXiv 2020

  8. [8]

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

Show all 45 references
  1. [9]

    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)

    2014 arXiv

  2. [10]

    Guy W Cole and Sinead A Williamson. 2019. Avoiding resentment via monotonic fairness. arXiv preprint arXiv:1909.01251 (2019)

    2019 arXiv

  3. [11]

    Hennie Daniels and Marina Velikova. 2010. Monotone and partially monotone neural networks. IEEE Transactions on Neural Networks 21, 6 (2010), 906–917

    2010

  4. [12]

    Ad J Feelders. 2000. Prior knowledge in economic applications of data mining. In European Conference on Principles of Data Mining and Knowledge Discovery . Springer, 395–400

    2000

  5. [13]

    Matt W Gardner and SR Dorling. 1998. Artificial neural networks (the multilayer perceptron)—a review of applications in the atmospheric sciences. Atmospheric environment 32, 14-15 (1998), 2627–2636

    1998

  6. [14]

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

    2017 arXiv

  7. [15]

    Akhil Gupta, Naman Shukla, Lavanya Marla, Arinbjörn Kolbeinsson, and Kartik Yellepeddi. 2019. How to incorporate monotonicity in deep networks while preserving flexibility? arXiv preprint arXiv:1909.10662 (2019)

    2019 arXiv

  8. [16]

    Malay Haldar, Mustafa Abdool, Prashant Ramanathan, Tao Xu, Shulin Yang, Huizhong Duan, Qing Zhang, Nick Barrow-Williams, Bradley C Turnbull, Bren- dan M Collins, et al. 2019. Applying deep learning to airbnb search. InProceedings of the 25th ACM SIGKDD International Conference...

    2019

  9. [17]

    Malay Haldar, Prashant Ramanathan, Tyler Sax, Mustafa Abdool, Lanbo Zhang, Aamir Mansawala, Shulin Yang, Bradley Turnbull, and Junshuo Liao. 2020. Im- proving deep learning for airbnb search. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery...

    2020

  10. [18]

    Xiangnan He, Lizi Liao, Hanwang Zhang, Liqiang Nie, Xia Hu, and Tat-Seng Chua. 2017. Neural collaborative filtering. In Proceedings of the 26th international conference on world wide web . 173–182

    2017

  11. [19]

    Zai Huang, Mingyuan Tao, and Bufeng Zhang. 2021. Deep Inclusion Relation- aware Network for User Response Prediction at Fliggy. In Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining . 3059–3067

    2021

  12. [20]

    Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. 1991. Adaptive mixtures of local experts. Neural computation 3, 1 (1991), 79–87

    1991

  13. [21]

    Xingchao Liu, Xing Han, Na Zhang, and Qiang Liu. 2020. Certified monotonic neural networks. Advances in Neural Information Processing Systems 33 (2020), 15427–15438

    2020

  14. [22]

    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

  15. [23]

    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. In The 41st International ACM SIGIR Conference on Research & Development in Information Retrie...

    2018

  16. [24]

    Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. 2013. Distributed representations of words and phrases and their compositionality. Advances in neural information processing systems 26 (2013)

    2013

  17. [25]

    An-phi Nguyen and María Rodríguez Martínez. 2019. Mononet: towards inter- pretable models by learning monotonic features. arXiv preprint arXiv:1909.13611 (2019)

    2019 arXiv

  18. [26]

    Yanru Qu, Han Cai, Kan Ren, Weinan Zhang, Yong Yu, Ying Wen, and Jun Wang

  19. [27]

    Steffen Rendle. 2010. Factorization machines. In2010 IEEE International conference on data mining. IEEE, 995–1000

    2010

  20. [28]

    In 2016 IEEE 16th International Conference on Data Mining (ICDM)

    Product-based neural networks for user response prediction. In 2016 IEEE 16th International Conference on Data Mining (ICDM) . IEEE, 1149–1154

    2016

  21. [29]

    Sculley, Gary Holt, Daniel Golovin, Eugene Davydov, Todd Phillips, Diet- mar Ebner, Vinay Chaudhary, Michael Young, Jean-François Crespo, and Dan Dennison

    D. Sculley, Gary Holt, Daniel Golovin, Eugene Davydov, Todd Phillips, Diet- mar Ebner, Vinay Chaudhary, Michael Young, Jean-François Crespo, and Dan Dennison. 2015. Hidden Technical Debt in Machine Learning Systems. In Ad- vances in Neural Information Processing Systems 28: An...

    2015

  22. [30]

    Steffen Rendle, Christoph Freudenthaler, Zeno Gantner, and Lars Schmidt-Thieme

  23. [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. In Fourteenth ACM Conference on Recommender Systems . 269– 278

    2020

  24. [33]

    David Simchi-Levi, Rui Sun, Michelle Wu, and Ruihao Zhu. 2022. Calibrating Sales Forecast in a Pandemic Using Competitive Online Non-Parametric Regression. In SSRN 3670264

    2022

  25. [34]

    Hong Wen, Jing Zhang, Yuan Wang, Fuyu Lv, Wentian Bao, Quan Lin, and Keping Yang. 2020. Entire space multi-task modeling via post-click behavior decomposi- tion for conversion rate prediction. In Proceedings of the 43rd International ACM SIGIR conference on research and develo...

    2020

  26. [35]

    Han Wu, Hongzhe Zhang, Liangyue Li, Zulong Chen, Fanwei Zhu, and Xiao Fang

  27. [36]

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

    2017

  28. [37]

    Jia Xu, Fei Xiong, Zulong Chen, Mingyuan Tao, Liangyue Li, and Quan Lu

  29. [38]

    Seungil You, David Ding, Kevin Canini, Jan Pfeifer, and Maya R. Gupta. 2017. Deep Lattice Networks and Partial Monotonic Functions. In Proceedings of the 31st International Conference on Neural Information Processing Systems (NIPS’17) . 2985–2993

    2017

  30. [39]

    Xiaokun Zhang, Bo Xu, Liang Yang, Chenliang Li, Fenglong Ma, Haifeng Liu, and Hongfei Lin. 2022. Price DOES Matter! Modeling Price and Interest Preferences in Session-based Recommendation. arXiv preprint arXiv:2205.04181 (2022)

    2022 arXiv

  31. [40]

    Kailun Wu, Weijie Bian, Zhangming Chan, Lejian Ren, Shiming Xiang, Shu- Guang Han, Hongbo Deng, and Bo Zheng. 2022. Adversarial Gradient Driven Exploration for Deep Click-Through Rate Prediction (KDD ’22). 2050–2058

    2022

  32. [41]

    Guorui Zhou, Xiaoqiang Zhu, Chenru Song, Ying Fan, Han Zhu, Xiao Ma, Yanghui Yan, Junqi Jin, Han Li, and Kun Gai. 2018. Deep interest network for click-through rate prediction. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining ...

    2018

  33. [42]

    In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining

    G2NET: A General Geography-Aware Representation Network for Hotel Search Ranking. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 4237–4247

  34. [45]

    Guorui Zhou, Na Mou, Ying Fan, Qi Pi, Weijie Bian, Chang Zhou, Xiaoqiang Zhu, and Kun Gai. 2019. Deep interest evolution network for click-through rate prediction. In Proceedings of the AAAI conference on artificial intelligence , Vol. 33. 5941–5948

    2019

  35. [2012]

    arXiv preprint arXiv:1205.2618 (2012)

    BPR: Bayesian personalized ranking from implicit feedback. arXiv preprint arXiv:1205.2618 (2012)

    2012 arXiv

  36. [2016]

    In Proceedings of the 1st workshop on deep learning for recommender systems

    Wide & deep learning for recommender systems. In Proceedings of the 1st workshop on deep learning for recommender systems . 7–10

  37. [2022]

    In 2022 IEEE 38th International Conference on Data Engineering (ICDE)

    Cheaper Is Better: Exploring Price Competitiveness for Online Purchase Prediction. In 2022 IEEE 38th International Conference on Data Engineering (ICDE) . IEEE, 3399–3412

    2022

Pith tools

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