Pith. sign in

REVIEW 3 major objections 6 minor 33 references

UniMVT jointly debiases CTR and estimates per-user coupon uplift in a single model

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

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

UniMVT jointly models debiased base CTR and coupon-induced uplift under a linear dose-response assumption and reports offline and online gains for coupon allocation.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection UniMVT is a plausible industrial recipe for joint debiased CTR and per-unit coupon uplift, but its convergence theorem is proven for a different estimator than the one the paper actually trains and serves. the 3 major comments →

arxiv 2602.12972 v2 pith:4UNXMXAY submitted 2026-02-13 cs.SI cs.LG

Jointly Optimizing Debiased CTR and Uplift for Coupons Marketing: A Unified Causal Framework

Siyun Yang , Shixiao Yang , Jian Wang , Di Fan , Kehe Cai , Haoyan Fu , Jiaming Zhang , Wenjin Wu
show 1 more author
Peng Jiang
This is my paper
classification cs.SI cs.LG
keywords multi-valued treatmentdebiased CTR predictionuplift modelingcausal inferenceunit upliftcoupon allocationmixture-of-expertscounterfactual inference
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

This paper claims that a single model can jointly produce a debiased click-through rate—the probability a user clicks without any coupon—and a per-user 'unit uplift' measuring how much each monetary unit of coupon increases that probability. The key move is to treat observed clicks during coupon campaigns as the sum of a baseline preference and a treatment effect, then train on the full population of treated and untreated users with counterfactual losses, so the two components are pinned down. The paper further assumes the treatment effect grows linearly with coupon value, which lets it compare coupons of any size by a single sensitivity number and allocate coupons by expected return on investment. If this works, platforms can stop letting coupons corrupt their CTR estimates and instead use the same model to decide who gets a coupon and how large it should be.

Core claim

The paper's central claim is that when the intensity-response relationship is monotonic and linear (Assumption 3.2), minimizing counterfactual calibration losses over the entire treated and control population makes the estimated unit uplift converge in probability to the true unit uplift (Theorem 4.1), while simultaneously yielding a base CTR estimate uncontaminated by coupon confounding. UniMVT operationalizes this with a mixture-of-experts encoder that separates treatment-invariant confounders from treatment-sensitive representations, two towers that estimate base and treated CTR, and an X-network that adds or subtracts the estimated uplift in logit space. The result is a single model whos

What carries the argument

The central identity is the decomposition of the conditional average treatment effect into the product of coupon intensity and a unit uplift: τ(x,t) = t·η(x), where η(x)>0. This reduces a continuous dose-response curve to one per-user scalar, making the uplift comparable across coupon values. The mechanism that carries the argument is the counterfactual X-network: predicted base CTR plus estimated uplift in logit space reconstructs the treated outcome, and predicted treated CTR minus the uplift reconstructs the baseline; the resulting mutual losses, together with a ReLU on η, enforce the monotonic linear structure and drive the convergence theorem.

Load-bearing premise

The assumption that a user's click-probability gain is exactly proportional to coupon value with no floor or saturation—so that the same per-unit sensitivity applies to every coupon size—is the load-bearing premise; if real response curves bend, both the debiased base CTR and the unit uplift are biased at deployment.

What would settle it

Run a randomized experiment where a segment of users receives coupons of value 10%, 20%, and 30%, and measure realized click-through gains. If the gain from 20% to 30% is materially smaller than the gain from 10% to 20% for many users, the linear monotonicity assumption fails and UniMVT's unit-uplift estimate will misrank coupon sizes. A cheaper check: train UniMVT on synthetic data generated from a concave dose-response curve and observe the drop in CS-AUUC on held-out RCT samples.

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

If this is right

  • Base CTR estimates from UniMVT are no longer inflated by coupon exposure, so ranking, pricing, and budget decisions can rely on a stable calibration.
  • Coupon size can be chosen per user by maximizing predicted revenue lift minus cost, with the option to withhold coupons whose ROI is below a threshold.
  • Training over the full click sample space, treated and untreated, exploits much more data than uplift models restricted to coupon recipients.
  • The unit-uplift formulation gives a single rankable sensitivity score across different coupon denominations, simplifying deployment.
  • If the linear assumption holds, the framework extends naturally to any continuous incentive, not just coupons.

Where Pith is reading between the lines

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

  • The linear monotonic assumption is the main point to check: real coupon response often saturates, and under a concave dose-response the estimated unit uplift would overstate the value of large coupons; the paper's own conclusion flags this as future work.
  • The framework could be extended to vector-valued unit effects for multiple simultaneous interventions (e.g., coupon plus ad format) by generalizing τ(x,t)=t·η(x) to a dot product.
  • The online ROI threshold γ is a policy lever that the model does not optimize; a separate tuning exercise against business constraints would complement the paper's contribution.
  • A direct validation of Assumption 3.2 would be to run stratified randomized experiments per user segment and check whether realized click-rate gains are proportional to coupon value across segments.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes UniMVT, a unified network for coupon-affected CTR prediction and uplift estimation under multi-valued treatments. The architecture uses a Mixture-of-Experts disentangled representation layer, separate base/treatment towers, an auxiliary intensity prediction head, and counterfactual losses that are meant to reconstruct base and treated CTR by adding or subtracting an estimated unit uplift. The authors claim that the model simultaneously yields debiased base CTR and precise unit uplift, supported by a convergence theorem, and they report results on synthetic, industrial, and online A/B data.

Significance. If the claims hold, the paper addresses an important practical problem: a single CTR model that is not biased by coupon interventions and provides per-user unit uplift for ROI-based coupon allocation. The industrial A/B results and the full-space modeling idea are potentially valuable, and the paper is careful to evaluate on synthetic data with ground truth and on a production dataset. However, the formal contribution is currently not supported: the counterfactual loss actually optimized differs from the loss analyzed in Theorem 4.1, and the proof itself has gaps. The empirical findings remain meaningful, but the theoretical guarantee, which is advertised as a main contribution, needs substantial revision.

major comments (3)
  1. [§4.3.2 (Eqs. 16–19), §4.4, §4.5 (Eq. 25), Appendix A] The training objective and the analyzed objective are different. Eqs. 16–17 define counterfactual predictions in logit space: p'_t = sigmoid(logit(p0_hat) + t_hat·eta_hat) and p'_0 = sigmoid(logit(p_t_hat) - t_hat·eta_hat), and these are used in Lx-treat/Lx-base (Eqs. 18–19). However, the proof of Theorem 4.1 (Eq. 22) and the appendix (Eqs. 28–36) replace p'_t by the additive probability form p0_hat + eta_hat·t_hat, and Eq. 25 also uses the additive form at inference. This is not a notational slip: the minimizer of the logit-space loss is not generally the eta satisfying p_t = p0 + eta·t, because sigmoid is nonlinear. Even under Assumptions 4.1–4.2, Theorem 4.1 does not establish convergence of the eta actually estimated by the model. The authors must either change Eqs. 16–17 (and the corresponding losses) to the additive form, or redo the theorem and inference in a consistent logit/prob
  2. [§4.4, Assumption 4.2, Appendix A.1] The proof of Theorem 4.1 uses assumptions that are not stated. Assumption 4.2 only postulates consistency of p_t_hat and t_hat, but the proof immediately uses |p0_hat - p0| → 0 (Eqs. 23–24) and an unstated Lipschitz bound on p_t with respect to t (Eq. 41). More fundamentally, the proof analyses the pointwise residual of a single sample and asserts that minimizing the joint loss drives that residual to zero; it never establishes consistency of the empirical risk minimizer or that the global minimizer of the joint loss attains vanishing residual. The theorem therefore overclaims: a triangle-inequality bound on a pointwise error is not a convergence theorem.
  3. [Theorem 4.1 / Definition 3.2] The unit uplift η is, by construction, the fitted quantity that explains the difference between the treated and baseline surfaces: p_t = p0 + t·η (Definition 3.2), and the counterfactual losses enforce exactly this relation. The theorem is close to a restatement of the modeling assumption; it does not provide an identifiability argument or a statistical convergence result for a joint estimator of (p0, t, η). In its current form, the claim that 'minimizing Lx-treat and Lx-base ensures η_hat converges to η' is not demonstrated beyond the algebra of the assumed linear model. A proper treatment should state the function classes, the loss landscape, and the conditions under which the empirical minimizer is consistent.
minor comments (6)
  1. [Figure 2 caption] The caption refers to 'CENIF' while the paper is titled UniMVT; this appears to be a leftover from an earlier draft and should be corrected.
  2. [Assumption 3.2, footnote 1] The superscript '1' after 'randomized experiments' has no corresponding footnote in the text; the empirical motivation for the monotonic linear relationship is therefore not documented.
  3. [§4.5 and Figure 4] The online pipeline labels both Base CTR and Unit Uplift as logit quantities, but Eq. 25 uses a probability-space additive form. The relationship between the logit-space training equations and the probability-space inference rule should be clarified.
  4. [§4.4 proof sketch] The paragraph before Eq. (22) says it takes Lx-treat as an example, but then states 'the counterfactual calibration loss for the control group'; this conflation of treated and control groups is confusing.
  5. [Throughout] There are typos and formatting inconsistencies, e.g., 'Futhermore' in §5.2 and inconsistent capitalization of 'CS-QINI' vs. 'CS-Qini'. The paper would benefit from a careful proofread.
  6. [§5.2 / Table 2] The synthetic results are reported without standard deviations or multiple seeds, and the synthetic data generation assumes exactly the linear monotonic response of Assumption 3.2; this makes the synthetic evaluation favorable to UniMVT and should be acknowledged.

Circularity Check

1 steps flagged

Theorem 4.1's proof makes convergence tautological by defining the minimized objective as the uplift error |η̂−η|, and it analyzes an additive-probability loss that the logit-space objective of Eqs. 16–19 does not optimize.

specific steps
  1. self definitional [Section 4.4, Proof Sketch of Theorem 4.1 (Eqs. 22–24)]
    "Let Δ=|η̂−η| be the objective to be minimized. Substituting the true p_t(x_i,t_i), we analyze the error term: Lx-treat =|(p̂0(x_i)+η̂ t̂)−(p0(x_i)+η·t)|"

    Theorem 4.1's conclusion is η̂→η. The proof begins by declaring the objective to be minimized to be Δ=|η̂−η|, which is exactly the quantity whose convergence the theorem is supposed to establish. The subsequent triangle inequality only re-expresses this definitional error. Furthermore, the proof replaces the actual counterfactual loss of Eq. 18—|y_i−σ(σ⁻¹(p̂0)+t̂η̂)|²—with the additive structural equation p̂0+η̂t̂. Thus the argument is not about the estimator actually trained; the claimed convergence is imposed by the proof's choice of objective rather than derived from the optimization problem.

full rationale

The circularity is localized to the paper's formal convergence claim. Definition 3.2 defines unit CATE as η=(p_t−p_0)/t, and the loss in Eqs. 16–19 fits η̂ to explain the residual between treated outcomes and the baseline. That by itself is standard supervised uplift estimation, not necessarily circular. However, Theorem 4.1—the paper's stated theoretical justification for 'precise uplift estimation'—proves convergence by setting the objective to Δ=|η̂−η| and by substituting the definitional additive relation for the actual logit-composed loss. This makes the theorem's central conclusion true by construction rather than as a consequence of the training procedure. The paper also exhibits a logit/additive mismatch: training uses p̂′_t=σ(σ⁻¹(p̂0)+t̂η̂) while the proof and the inference rule of Eq. 25 use p̂0+η̂q, so even a non-circular reading would not establish convergence of the implemented estimator. That is a correctness gap supporting the score. No load-bearing self-citation is present: the cited representation, meta-learner, and continuous-treatment works are external, and the empirical results on synthetic, industrial, and online A/B data provide independent evidence for the method's practical behavior. The self-reported limitation in the conclusion—'future work will explore modeling techniques that better align with complex intensity-response curves'—and the missing footnote 1 are noted but are not additional circularity.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The central claim rests on two structural assumptions beyond standard causal identifiability: linear uplift in dose (Assumption 3.2) and cross-population support/consistency for intensity imputation (Assumption 4.1). The convergence theorem adds a consistency assumption that largely assumes its own conclusion. No new physical or causal entities are introduced.

free parameters (4)
  • Loss weights λ_base, λ_treat, λ_t, λ_X, λ_o = not reported
    Eq. (20) balances five loss terms; no values or sensitivity analysis are provided, so the final objective is under-specified for replication.
  • Intensity-loss weights λ1, λ2 = not reported
    Eq. (14) uses λ1 and λ2 for L1/L2 intensity prediction; values are not given.
  • Intensity range [t_min, t_max] = not reported
    Eq. (13) projects predicted intensity into [t_min, t_max]; these bounds are data-dependent and not specified.
  • Architecture hyperparameters = not reported
    Number of experts, hidden layer sizes, learning rate, and optimizer settings are not reported in the manuscript.
axioms (5)
  • standard math Consistency, unconfoundedness, and overlap (Assumption 3.1)
    Standard potential-outcomes assumptions [20] needed for CATE identifiability; they are not validated on the industrial data but are routinely invoked in this literature.
  • domain assumption Monotonic linear relationship: τ(x,t) ≈ α(x)·t with α(x) > 0 (Assumption 3.2)
    Load-bearing. The entire architecture (ReLU on η, counterfactual losses, inference rule Eq. 25) assumes uplift is linear in coupon value with zero intercept. The paper says it is motivated by randomized experiments but the footnote content and citations are missing.
  • ad hoc to paper Non-trivial intensity and support inclusion (Assumption 4.1)
    Needed for the proof of Theorem 4.1: |t| ≥ t_min > 0 and the support of control-group intensities is contained in the treated-group support. This is an identification assumption, not empirically checked.
  • ad hoc to paper Consistency of base estimators p_t and t (Assumption 4.2)
    Theorem 4.1 assumes that p_t and t are already consistent, which does most of the work; the theorem then 'concludes' η is consistent. This is a near-tautological assumption rather than an established result.
  • ad hoc to paper Lipschitz continuity of the true response p_t with respect to t
    Appendix A.1.2 uses a bound ||p_t(x,t_hat) - p_t(x,t)|| ≤ L_p ||t_hat - t|| to show the perturbation error vanishes; this Lipschitz condition is never stated or justified.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Jointly Optimizing Debiased CTR and Uplift for Coupons Marketing: A Unified Causal Framework." pith.science (2026). https://pith.science/paper/4UNXMXAY

@misc{pith2026260212972,
  author       = {Pith},
  title        = {Pith review of: Jointly Optimizing Debiased CTR and Uplift for Coupons Marketing: A Unified Causal Framework},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4UNXMXAY}},
  note         = {Machine review of arXiv:2602.12972}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

In online advertising, marketing interventions such as coupons introduce significant confounding bias into Click-Through Rate (CTR) prediction. Observed clicks reflect a mixture of users' intrinsic preferences and the uplift induced by these interventions. This causes conventional models to miscalibrate base CTRs, which distorts downstream ranking and billing decisions. Furthermore, marketing interventions often operate as multi-valued treatments with varying magnitudes, introducing additional complexity to CTR prediction. To address these issues, we propose the \textbf{Uni}fied \textbf{M}ulti-\textbf{V}alued \textbf{T}reatment Network (UniMVT). Specifically, UniMVT disentangles confounding factors from treatment-sensitive representations, enabling a full-space counterfactual inference module to jointly reconstruct the debiased base CTR and intensity-response curves. To handle the complexity of multi-valued treatments, UniMVT employs an auxiliary intensity estimation task to capture treatment propensities and devise a unit uplift objective that normalizes the intervention effect. This ensures comparable estimation across the continuous coupon-value spectrum. UniMVT simultaneously achieves debiased CTR prediction for accurate system calibration and precise uplift estimation for incentive allocation. Extensive experiments on synthetic and industrial datasets demonstrate UniMVT's superiority in both predictive accuracy and calibration. Furthermore, real-world A/B tests confirm that UniMVT significantly improves business metrics through more effective coupon distribution.

Figures

Figures reproduced from arXiv: 2602.12972 by Di Fan, Haoyan Fu, Jiaming Zhang, Jian Wang, Kehe Cai, Peng Jiang, Shixiao Yang, Siyun Yang, Wenjin Wu.

Figure 1
Figure 1. Figure 1: Motivation analysis of causal marketing interventions and sensitivity estimation. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of CENIF. The framework integrates two core modules: (i) a Deconfounded Causal Representation (DCR) [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: CS-AUUC comparison of different methods across [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Illustration of the online pipeline [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The illustration of coupon scenario. These pre-click [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

33 extracted references · 1 linked inside Pith

  1. [1]

    Meng Ai, Zhuo Chen, Jibin Wang, Jing Shang, Tao Tao, and Zhen Li. 2024. Improve roi with causal learning and conformal prediction. In2024 IEEE 40th International Conference on Data Engineering (ICDE). IEEE, 598–610

    2024

  2. [2]

    Qiwei Bi, Jian Li, Lifeng Shang, Xin Jiang, Qun Liu, and Hanfang Yang. 2022. Mtrec: Multi-task learning over bert for news recommendation. InFindings of the association for computational linguistics: ACL 2022. 2663–2669

    2022

  3. [3]

    Rich Caruana. 1997. Multitask learning.Machine learning28 (1997), 41–75

    1997

  4. [4]

    Jianxin Chang, Chenbin Zhang, Zhiyi Fu, Xiaoxue Zang, Lin Guan, Jing Lu, Yiqun Hui, Dewei Leng, Yanan Niu, Yang Song, et al. 2023. TWIN: TWo-stage interest network for lifelong user behavior modeling in CTR prediction at kuaishou. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 3785–3794

    2023

  5. [5]

    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(Long Beach, CA, USA)(KDD ’23). Association for Computing Machinery, New Y...

    2023

  6. [6]

    Alicia Curth and Mihaela Van der Schaar. 2021. On inductive biases for hetero- geneous treatment effect estimation.Advances in Neural Information Processing Systems34 (2021), 15883–15894

    2021

  7. [7]

    Zichuan Fu, Xiangyang Li, Chuhan Wu, Yichao Wang, Kuicai Dong, Xiangyu Zhao, Mengchen Zhao, Huifeng Guo, and Ruiming Tang. 2025. A unified frame- work for multi-domain ctr prediction via large language models.ACM Transac- tions on Information Systems43, 5 (2025), 1–33

    2025

  8. [8]

    Huifeng Guo, TANG Ruiming, Yunming Ye, Zhenguo Li, and Xiuqiang He. 2017. DeepFM: A Factorization-Machine based Neural Network for CTR Prediction. In Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelli- gence. International Joint Conferences on Artificial Intelligence Organization

    2017

  9. [9]

    Bowei He, Yunpeng Weng, Xing Tang, Ziqiang Cui, Zexu Sun, Liang Chen, Xi- uqiang He, and Chen Ma. 2024. Rankability-enhanced revenue uplift modeling framework for online marketing. InProceedings of the 30th ACM SIGKDD Confer- ence on Knowledge Discovery and Data Mining. 5093–5104

    2024

  10. [10]

    Shen Jiang, Guanghui Zhu, Yue Wang, Chunfeng Yuan, and Yihua Huang. 2024. Automatic Multi-Task Learning Framework with Neural Architecture Search in Recommendations. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 1290–1300

    2024

  11. [11]

    Wenwei Ke, Chuanren Liu, Xiangfu Shi, Yiqiao Dai, Philip S Yu, and Xiaoqiang Zhu. 2021. Addressing exposure bias in uplift modeling for large-scale online advertising. In2021 IEEE International Conference on Data Mining (ICDM). IEEE, 1156–1161

    2021

  12. [12]

    Sören R Künzel, Jasjeet S Sekhon, Peter J Bickel, and Bin Yu. 2019. Metalearners for estimating heterogeneous treatment effects using machine learning.Proceedings of the national academy of sciences116, 10 (2019), 4156–4165

    2019

  13. [13]

    Danwei Li, Zhengyu Zhang, Siyang Yuan, Mingze Gao, Weilin Zhang, Chaofei Yang, Xi Liu, and Jiyan Yang. 2023. Adatt: Adaptive task-to-task fusion network for multitask learning in recommendations. InProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 4370–4379

    2023

  14. [14]

    Haoxuan Li, Kunhan Wu, Chunyuan Zheng, Yanghao Xiao, Hao Wang, Zhi Geng, Fuli Feng, Xiangnan He, and Peng Wu. 2023. Removing hidden confounding in recommendation: a unified multi-task learning approach.Advances in Neural Information Processing Systems36 (2023), 54614–54626

    2023

  15. [15]

    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

  16. [16]

    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

    2018

  17. [17]

    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 Retrieval. 1137–1140

    2018

  18. [18]

    Kelong Mao, Jieming Zhu, Liangcai Su, Guohao Cai, Yuru Li, and Zhenhua Dong

  19. [19]

    Lizhen Nie, Mao Ye, Qiang Liu, and Dan Nicolae. 2021. Vcnet and functional targeted regularization for learning causal effects of continuous treatments.arXiv preprint arXiv:2103.07861(2021)

    Pith/arXiv arXiv 2021

  20. [20]

    Donald B Rubin. 2005. Causal Inference Using Potential Outcomes.J. Amer. Statist. Assoc.100, 469 (2005), 322–331

    2005

  21. [21]

    Patrick Schwab, Lorenz Linhardt, Stefan Bauer, Joachim M Buhmann, and Walter Karlen. 2020. Learning counterfactual representations for estimating individ- ual dose-response curves. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 34. 5612–5619

    2020

  22. [22]

    Uri Shalit, Fredrik D Johansson, and David Sontag. 2017. Estimating individual treatment effect: generalization bounds and algorithms. InInternational conference on machine learning. PMLR, 3076–3085

    2017

  23. [23]

    Liangcai Su, Junwei Pan, Ximei Wang, Xi Xiao, Shijie Quan, Xihua Chen, and Jie Jiang. 2024. STEM: unleashing the power of embeddings for multi-task recommendation. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 38. 9002–9010

    2024

  24. [24]

    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. [25]

    Wanjie Tao, Huihui Liu, Xuqi Li, Qun Dai, Hong Wen, and Zulong Chen. 2023. Event-Aware Adaptive Clustering Uplift Network for Insurance Creative Ranking. InProceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1966–1970

    2023

  26. [26]

    Fan Wang, Lianyong Qi, Weiming Liu, Bowen Yu, Jintao Chen, and Yanwei Xu. 2025. Inter-and intra-similarity preserved counterfactual incentive effect estimation for recommendation systems.ACM Transactions on Information Systems43, 6 (2025), 1–24

    2025

  27. [27]

    Fangye Wang, Yingxu Wang, Dongsheng Li, Hansu Gu, Tun Lu, Peng Zhang, and Ning Gu. 2022. Enhancing CTR prediction with context-aware feature repre- sentation learning. InProceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval. 343–352

    2022

  28. [28]

    Yejing Wang, Zhaocheng Du, Xiangyu Zhao, Bo Chen, Huifeng Guo, Ruiming Tang, and Zhenhua Dong. 2023. Single-shot feature selection for multi-task recommendations. InProceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval. 341–351

    2023

  29. [29]

    Haizhi Yang, Tengyun Wang, Xiaoli Tang, Qianyu Li, Yueyue Shi, Siyu Jiang, Han Yu, and Hengjie Song. 2021. Multi-task learning for bias-free joint ctr prediction and market price modeling in online advertising. InProceedings of the 30th ACM International Conference on Information & Knowledge Management. 2291–2300

    2021

  30. [30]

    Xin Zhang, Kai Wang, Zengmao Wang, Bo Du, Shiwei Zhao, Runze Wu, Xudong Shen, Tangjie Lv, and Changjie Fan. 2024. Temporal Uplift Modeling for Online Marketing. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 6247–6256

    2024

  31. [31]

    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 systems. 43–51

    2019

  32. [32]

    Kailiang Zhong, Fengtong Xiao, Yan Ren, Yaorong Liang, Wenqing Yao, Xiaofeng Yang, and Ling Cen. 2022. Descn: Deep entire space cross networks for individual treatment effect estimation. InProceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining. 4612–4620. A Theoretical Analysis A.1 Proof of Theorem 4.1 We aim to show that min...

    2022

  33. [2023]

    In Proceedings of the AAAI conference on artificial intelligence, Vol

    FinalMLP: an enhanced two-stream MLP model for CTR prediction. In Proceedings of the AAAI conference on artificial intelligence, Vol. 37. 4552–4560

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.