Pith. sign in

REVIEW 3 major objections 7 minor 49 references

CanniUplift: A Holistic Framework for Mitigating Seller and Incentive Cannibalization in E-commerce Uplift Modeling

T0 review · 3 major / 7 minor · reviewed 2026-07-07 · glm-5.2

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

Pith's one-line read E-commerce coupons cannibalize the platform they aim to grow

desk verdict Production uplift modeling paper with a real deployment and a genuine problem framing, but the central inference approximation is under-justified. read the letter →

arxiv 2607.05242 v1 pith:XF445AL3 submitted 2026-07-06 cs.LG cs.AIcs.IR

Zuwang He , Shihao Shu , Yuli Qu , Hanyu Gao , Ziliang Zhang , Diwei Chen , Xiangda Yan , Buyu Gao
show 3 more authors
Tanchao Zhu Yumeng Li Junxiong Zhu
This is my paper · ORCID
classification cs.LGcs.AIcs.IR
keywords upliftmodelingcannibalizationSUTVAviolationcausalinferencee-commerceindividualtreatmenteffectGMVoptimizationincentiveallocation
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 identifies two systematic ways that standard uplift models waste marketing budgets in multi-seller e-commerce platforms. First, a coupon for one seller may merely shift a user's spending away from another seller on the same platform, producing apparent seller-level lift but zero net platform growth (seller-level cannibalization). Second, a user who receives a coupon may purchase without redeeming it—driven by organic intent or a competing coupon—causing the model to falsely credit the assigned coupon for the conversion (incentive-level cannibalization). The authors propose CanniUplift, a framework that addresses both problems. To mitigate seller-level cannibalization, a Platform-level Global Alignment (PGA) module adds a training loss that constrains the sum of per-seller GMV predictions to match the user's actual total platform GMV, penalizing the model when spurious seller-level uplift inflates the aggregate. To mitigate incentive-level cannibalization, a Redemption-based Decomposition Denoising (RDD) module splits the treated-outcome prediction into two paths—purchase with redemption and purchase without redemption—so that non-redemption conversions are not forced to share the same uplift pattern as genuine coupon-driven conversions. A Treat-Attention mechanism generates candidate-specific representations by attending from each candidate seller/coupon to the user's historical behavior sequence. The framework was deployed on a large-scale production system, where online A/B testing showed a 4.08% relative increase in platform-wide incremental GMV and a 6.69% improvement in ROI compared to the production baseline.

What carries the argument

The load-bearing premise is the inference approximation in Appendix C: at serving time, the platform-level marginal uplift for a single seller is estimated by differencing two single-point platform-head predictions (one with treatment, one without), rather than recomputing counterfactual outcomes for all other sellers. The paper argues this is valid because the global aggregation loss during training has already encoded cross-seller dependencies into the candidate-specific representation. If that encoding is insufficient—if the training loss does not fully capture how treating one seller changes another seller's outcome—the single-point approximation yields biased marginal uplift estimates.

What would settle it

If the global aggregation loss during training does not sufficiently encode cross-seller dependencies into the candidate-specific representation, the single-point inference approximation yields biased marginal uplift estimates.

Watch

Extended reading notes

Core claim

The central mechanism is the combination of an additive global consistency constraint (PGA) with a redemption-path decomposition (RDD). PGA works by making the model's per-seller predictions sum to the observed platform total during training; if the model assigns high uplift to multiple competing sellers simultaneously, the inflated aggregate is penalized, which implicitly teaches the model to discount substitution effects without explicitly modeling every seller pair. RDD works by factorizing the treated outcome into a redemption-weighted mixture of two path-specific increments, separating genuine coupon-driven conversions from conversions that occur without redemption. The two modules are:

Load-bearing premise

If the global aggregation training loss sufficiently encodes cross-seller substitution into candidate-specific representations, then single-point inference is a valid and computationally cheap approximation of marginal platform uplift.

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 / 7 minor

Summary. The paper proposes CanniUplift, a framework for e-commerce uplift modeling that addresses two forms of cannibalization: seller-level (cross-shop substitution under SUTVA violation) and incentive-level (noncompliance/noise from unredeemed coupons). The framework introduces Platform-level Global Alignment (PGA), which trains a platform head under an additive GMV consistency constraint, and Redemption-based Decomposition Denoising (RDD), which decomposes treated outcomes into redemption and non-redemption paths. A Treat-Attention mechanism generates candidate-specific representations. The paper reports improvements over baselines on synthetic and industrial datasets (Table 1), ablation results confirming each module's contribution (Table 2), and a 4.08% relative increase in platform-wide incremental GMV in an online A/B test (Table 3).

Significance. The problem addressed is practically important: SUTVA violations in multi-seller e-commerce uplift modeling are real and underexplored. The dual-source cannibalization framing (seller-level and incentive-level) is a useful conceptual contribution. The online A/B test with production-scale deployment and positive ROI impact is a significant strength. The ablation design isolating RDD from mere auxiliary redemption prediction (Table 2, '+Redem' vs '+RDD') is well-constructed. The synthetic dataset with an explicit cannibalization decay mechanism (Appendix A.6) provides a controlled testbed for PGA. The Treat-Attention mechanism for candidate-specific representation is a reasonable design choice.

major comments (3)
  1. Appendix C.2–C.4, Eq. (33) vs. Eq. (14): The inference approximation is the load-bearing mechanism for PGA's deployment, and its theoretical justification is insufficient. The training loss (Eq. 14) enforces that the *sum* of per-seller platform-head predictions matches total observed platform GMV in the observed treatment configuration. At inference, the needed quantity is the *marginal* counterfactual effect of treating seller s_j on total platform GMV (Eq. 31), which includes cross-seller spillover terms E[y_{u,s_k}|t_j=1] - E[y_{u,s_k}|t_j=0] for s_k ≠ s_j. The paper states (Appendix C.3) that cross-seller substitution is 'implicitly captured through the parameters learned from the platform-level aggregation loss,' but an additive consistency constraint on predicted levels does not uniquely determine per-seller marginal effects. Multiple decompositions of total GMV into per-seller贡献s
  2. satisfy the sum constraint, and not all of them correctly encode the marginal platform uplift. The paper should either (a) provide a formal argument for why the training signal is sufficient to identify marginal effects, or (b) directly validate the approximation on the synthetic dataset, where ground-truth marginal platform uplift is computable. Specifically, on the synthetic data, the authors could compare the single-point approximation δ_platform(u, s_j) from Eq. (33) against the true ΔPlatform(u, s_j) from Eq. (31) and report the correlation or error. This would directly test the load-bearing premise.
  3. §4.6, Eq. (21): The cannibalization rate g = (p_seller - p_platform) / p_seller is defined entirely from two model outputs. The validation in Figure 3 shows that g correlates with behavioral features (same-category browsing, price ratios), but since g is mechanically derived from the model's own predictions, this is not independent validation that the model correctly estimates cannibalization. The paper should acknowledge this limitation more explicitly, or provide an external validation signal (e.g., comparing g against the ground-truth cannibalization rate in the synthetic dataset, where the decay mechanism in Eq. 29 provides a known cross-effect).
minor comments (7)
  1. §4.1.2: The synthetic dataset only simulates seller-level cannibalization, not incentive-level. While the paper acknowledges this and validates RDD on industrial data, a synthetic dataset with controlled incentive-level noise would strengthen the RDD validation. Consider adding this in future work.
  2. Table 1: The synthetic dataset results show 'Ours' as PGA added to EUEN, but it would be useful to also report the full model (PGA+RDD) on synthetic data to confirm that RDD does not degrade performance when incentive cannibalization is absent.
  3. §3.5, Eq. (16): The loss weights λ_pla and λ_redem are set to 1, but no sensitivity analysis is provided. Given that these control the strength of the global alignment and denoising signals, a brief ablation over these values would be informative.
  4. Figure 2: The architecture diagram is dense. Clarifying which arrows represent gradient flow during training vs. inference-time computation would improve readability.
  5. Appendix A.6, Eq. (29): The decay coefficient d = 4 + 2(n-1) when n ≥ 1 seems large relative to the uplift terms. It would help to state the typical magnitude of uplift_raw_i to contextualize the 50% cap.
  6. §4.8: The online A/B test runs for only 8 days (Oct 5–12, 2025). A longer test period would strengthen the claim of sustained improvement, particularly given the temporal cannibalization limitation acknowledged in §7.
  7. The paper uses both 'Head1' and 'Platform head' to refer to the same component. Standardizing terminology would improve clarity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for a careful and substantive review. Both major comments identify genuine gaps in our validation strategy. We address each below and commit to specific revisions.

read point-by-point responses

  1. Referee: Appendix C.2–C.4, Eq. (33) vs. Eq. (14): The inference approximation is the load-bearing mechanism for PGA's deployment, and its theoretical justification is insufficient. The training loss enforces that the sum of per-seller platform-head predictions matches total observed platform GMV in the observed treatment configuration. At inference, the needed quantity is the marginal counterfactual effect of treating seller s_j on total platform GMV (Eq. 31), which includes cross-seller spillover terms. An additive consistency constraint on predicted levels does not uniquely determine per-seller marginal effects. The paper should either (a) provide a formal argument for why the training signal is sufficient to identify marginal effects, or (b) directly validate the approximation on the synthetic dataset, where ground-truth marginal platform uplift is computable.

    Authors: The referee is correct that the sum constraint in Eq. (14) does not, by itself, formally identify per-seller marginal platform effects. Multiple decompositions of total platform GMV into per-seller contributions satisfy the additive consistency constraint, and not all of them correctly encode the marginal platform uplift ΔPlatform(u, s_j) as defined in Eq. (31). Our current manuscript overstates the case when it says cross-seller substitution is 'implicitly captured through the parameters learned from the platform-level aggregation loss' without providing either a formal identification argument or direct empirical validation of the approximation. We cannot offer a formal identification proof because, in general, the sum constraint alone is insufficient for identification — the referee's mathematical observation is accurate. What we can offer is the direct empirical validation the referee suggests in option (b). On the synthetic dataset, the ground-truth marginal platform uplift ΔPlatform(u, s_j) is computable: for each user–seller pair, we can compute the total platform GMV under t_j = 1 versus t_j = 0 by re-evaluating the generative process (including the decay mechanism in Eq. 29) across all sellers in S_u. We will run this comparison and report the correlation (and mean absolute error) between the single-point approximation δ_platform(u, s_j) from Eq. (33) and the true ΔPlatform(u, s_j) from Eq. (31). We will add these results to the revised manuscript, along with an explicit acknowledgment that the sum constraint provides a regularization signal rather than a formal identification guarantee, and that the practical validity of the approximation rests on the empirical validation we will now provide. We also note that the online A/B test results (Table 3) provide an间接但 revision: no

  2. Referee: §4.6, Eq. (21): The cannibalization rate g = (p_seller - p_platform) / p_seller is defined entirely from two model outputs. The validation in Figure 3 shows that g correlates with behavioral features, but since g is mechanically derived from the model's own predictions, this is not independent validation that the model correctly estimates cannibalization. The paper should acknowledge this limitation more explicitly, or provide an external validation signal (e.g., comparing g against the ground-truth cannibalization rate in the synthetic dataset, where the decay mechanism in Eq. 29 provides a known cross-effect).

    Authors: The referee is correct. The validation in Figure 3 is not independent of the model: g is computed from p_seller and p_platform, both of which are model outputs, so the correlation with behavioral features demonstrates internal consistency of the model's predictions with known cannibalization-associated behaviors, but does not establish that g accurately estimates the true cannibalization rate. We will revise the manuscript to explicitly state this limitation in §4.6, clarifying that Figure 3 provides a plausibility check rather than independent validation. More importantly, we will add the external validation the referee suggests. On the synthetic dataset, the decay mechanism in Eq. (29) provides a computable ground-truth cross-effect for each user–seller pair: the difference between the raw uplift and the decayed uplift captures the cannibalization amount attributable to prior same-segment exposure. We will compute a ground-truth cannibalization rate from this quantity and compare it against the model-estimated g, reporting correlation and calibration metrics. This will provide the independent validation signal that is currently missing. revision: no

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity found; the cannibalization rate diagnostic is derived from model outputs but is not a load-bearing prediction.

full rationale

The paper's central claims are validated against external benchmarks: wAUUC/wQINI on synthetic and industrial datasets (Table 1), ablation studies against baselines (Table 2), and online A/B test results measuring real platform ΔGMV and ROI (Table 3). These are not reducible to the model's own outputs by construction. The reader's concern about Eq. 21 (the cannibalization rate g = (p_seller - p_platform) / p_seller) is understandable but does not constitute circularity in the sense required here: the paper defines g as a diagnostic metric in §4.6 and checks whether it correlates with external behavioral features (same-category browsing counts, price ratios) in Fig. 3. The paper does not claim g is a first-principles prediction or a derived result; it is explicitly a post-hoc analysis tool ('To verify whether the model effectively learns cannibalization effects, we define...'). The correlation in Fig. 3 is not mechanically forced because the behavioral features are external to the model. The inference approximation in Appendix C.2–C.4 (Eq. 33) is a modeling assumption that the paper transparently acknowledges as an approximation ('This approximation should not be interpreted as assuming that other sellers' outcomes are unaffected'). Whether this approximation is valid is a correctness risk, not circularity — the training loss (Eq. 14) and inference procedure (Eq. 33) are different operations, and the gap between them is a limitation the paper discloses rather than hides. No self-citation chains are load-bearing: the paper cites standard causal inference literature (Rubin, Holland, Künzel, etc.) and related uplift methods, none of which are authored by the present paper's authors. The derivation chain from problem definition through architecture design to empirical validation is self-contained against external benchmarks.

Assumptions & free parameters 6 free parameters · 4 assumptions · 2 invented entities

The framework introduces 6 free parameters (2 loss weights, 1 Tweedie p, 2 Optuna-searched hyperparameters, 1 synthetic decay coefficient), 4 axioms (2 domain assumptions, 1 ad-hoc-to-paper, 1 standard math), and 2 invented entities (1 architectural component with partial independent evidence, 1 derived metric without independent evidence). The ad-hoc inference approximation axiom is the most concerning: it is load-bearing for the deployment claim but lacks formal justification.

free parameters (6)
  • λ_pla = 1.0
    Weight of platform-level aggregation loss, set to 1 'unless otherwise specified' (§3.5). Not tuned via Optuna despite Optuna being used for other hyperparameters.
  • λ_redem = 1.0
    Weight of redemption classification loss, set to 1 'unless otherwise specified' (§3.5). Same as above.
  • Tweedie distribution parameter p = Not specified (1 < p < 2)
    The Tweedie loss parameter p controls the distribution shape. The paper states 1 < p < 2 but does not report the value used. This is a free parameter that affects GMV fitting.
  • Learning rate η = Searched in {5e-4, 1e-4, 5e-5}
    Optuna search space (§4.3).
  • Hidden dimension d_h = Searched in {128, 256, 512}
    Optuna search space (§4.3).
  • Synthetic cannibalization decay coefficient d = 4 + 2(n-1) for n≥1, else 0
    Hand-set decay parameter in synthetic data generation (Appendix A.6, Eq. 29). Controls the magnitude of simulated cannibalization.
assumptions (4)
  • domain assumption SUTVA is violated in multi-seller e-commerce due to cross-seller substitution and incentive noncompliance
    Stated in §1 and §2.3. This is the foundational premise motivating the entire framework. Supported by the observation that baseline models overestimate total gain by ~35% on industrial data (§4.1.1).
  • ad hoc to paper The global aggregation loss during training sufficiently encodes cross-seller substitution effects into model parameters such that single-point inference approximates marginal platform uplift
    Appendix C.3. This is the key unstated assumption enabling practical deployment. No formal bound or empirical validation against ground-truth marginal effects is provided.
  • domain assumption Redemption behavior is a valid post-treatment variable for decomposing treated outcomes into incremental and non-incremental paths
    §2.4 and §3.4. The paper assumes that redemption vs. non-redemption cleanly separates coupon-driven from organic conversions. In reality, redemption can occur for organic conversions (user would have bought anyway but also uses the coupon), which is not addressed.
  • standard math The Tweedie distribution (1 < p < 2) is an appropriate model for zero-inflated, long-tailed GMV data
    §3.5. Standard distributional assumption for compound Poisson-gamma data, reasonable for GMV modeling.
invented entities (2)
  • Platform head (Head1) independent evidence
    purpose: Aggregates per-seller predictions and aligns with total platform GMV via consistency loss
    The platform head is a new architectural component whose predictions are validated against observed platform GMV (Eq. 9) and whose cannibalization estimates correlate with behavioral features (Fig. 3). However, the inference approximation it enables (single-point marginal scoring) is not independently validated against ground-truth counterfactuals.
  • Cannibalization rate g (Eq. 21)
    purpose: Derived metric quantifying the gap between seller-level and platform-level predicted uplift
    Defined as (p_seller - p_platform) / p_seller. This is a derived quantity from two model outputs, not an independently measured entity. Its correlation with behavioral features (Fig. 3) is suggestive but not independent evidence since the model was trained on those same features.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CanniUplift: A Holistic Framework for Mitigating Seller and Incentive Cannibalization in E-commerce Uplift Modeling." pith.science (2026). https://pith.science/paper/XF445AL3

@misc{pith2026260705242,
  author       = {Pith},
  title        = {Pith review of: CanniUplift: A Holistic Framework for Mitigating Seller and Incentive Cannibalization in E-commerce Uplift Modeling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XF445AL3}},
  note         = {Machine review of arXiv:2607.05242}
}
read the original abstract

Personalized incentive allocation is vital for e-commerce, where uplift modeling is the standard for estimating Individual Treatment Effects (ITE). However, traditional models often fail in complex multi-seller environments with violations of the Stable Unit Treatment Value Assumption (SUTVA). We identify two critical challenges: Seller-level Cannibalization, where incentives shift expenditure between shops without growing the platform, and Incentive-level Cannibalization, where organic conversions or alternative rewards introduce significant noise into incrementality estimation. In this paper, we propose CanniUplift, a unified framework to mitigate these dual-source cannibalization effects. Specifically, we design Platform-level Global Alignment (PGA) to capture cross-shop substitution through global GMV consistency constraints. To tackle incentive-driven noise, we introduce Redemption-based Decomposition Denoising (RDD), which uses redemption behavior to decompose treated outcomes and reduce attribution noise within an entire-space framework. Furthermore, a Treat-Attention mechanism is designed to model intricate interactions between users' historical behaviors and current treatment options. Extensive experiments on both synthetic and large-scale industrial datasets demonstrate that CanniUplift significantly outperforms state-of-the-art baselines. Ablation studies confirm that the integration of PGA and RDD consistently improves wAUUC and wQINI. Successfully deployed online, our framework achieved a 4.08% relative increase in platform-wide incremental GMV (Delta GMV) over the production baseline and improved ROI in online A/B tests, proving effective in driving global platform growth.

Figures

Figures reproduced from arXiv: 2607.05242 by the authors.

Figure 1
Figure 1. Two critical violations of SUTVA in multi-seller [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The architecture of the CanniUplift framework. The model consists of three core components: (Left) The Platform-level [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Correlation between estimated cannibalization rate [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

49 extracted references · 49 canonical work pages

  1. [1]

    Naoufal Acharki, Ramiro Lugo, Antoine Bertoncello, and Josselin Garnier. 2023. Comparison of meta-learners for estimating multi-valued treatment heteroge- neous effects. InInternational conference on machine learning. PMLR, 91–132

    work page 2023

  2. [2]

    Takuya Akiba, Shotaro Sano, Toshihiko Yanase, Takeru Ohta, and Masanori Koyama. 2019. Optuna: A next-generation hyperparameter optimization frame- work. InProceedings of the 25th ACM SIGKDD international conference on knowl- edge discovery & data mining. 2623–2631

    work page 2019

  3. [3]

    Serge Assaad, Shuxi Zeng, Chenyang Tao, Shounak Datta, Nikhil Mehta, Ricardo Henao, Fan Li, and Lawrence Carin. 2021. Counterfactual representation learning with balancing weights. InInternational Conference on Artificial Intelligence and Statistics. PMLR, 1972–1980

    work page 2021

  4. [4]

    Susan Athey and Guido Imbens. 2016. Recursive partitioning for heterogeneous causal effects.Proceedings of the National Academy of Sciences113, 27 (2016), 7353–7360

    work page 2016

  5. [5]

    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

    work page 2021

  6. [6]

    Zhihao Guo, Peng Song, Chenjiao Feng, Kaixuan Yao, Chuangyin Dang, and Jiye Liang. 2025. Causal intervention for knowledge graph denoising in recommender KDD ’26, August 09–13, 2026, Jeju Island, Republic of Korea Zuwang He et al. systems.International Journal of Machine Learning and Cybernetics16, 11 (2025), 8551–8567

    work page 2025

  7. [7]

    Pierre Gutierrez and Jean-Yves Gérardy. 2017. Causal inference and uplift mod- elling: A review of the literature. InInternational conference on predictive applica- tions and APIs. PMLR, 1–13

    work page 2017

  8. [8]

    Behram Hansotia and Brad Rukstales. 2002. Incremental value modeling.Journal of Interactive Marketing16, 3 (2002), 35–46

    work page 2002

Show all 49 references
  1. [9]

    Paul W Holland. 1986. Statistics and causal inference.Journal of the American statistical Association81, 396 (1986), 945–960

    1986

  2. [10]

    2015.Causal inference in statistics, social, and biomedical sciences

    Guido W Imbens and Donald B Rubin. 2015.Causal inference in statistics, social, and biomedical sciences. Cambridge university press

    2015

  3. [11]

    Jiarui Jin, Xianyu Chen, Weinan Zhang, Yuanbo Chen, Zaifan Jiang, Zekun Zhu, Zhewen Su, and Yong Yu. 2022. Multi-scale user behavior network for entire space multi-task learning. InProceedings of the 31st ACM International Conference on Information & Knowledge Management. 874–883

    2022

  4. [12]

    Fredrik Johansson, Uri Shalit, and David Sontag. 2016. Learning representations for counterfactual inference. InInternational conference on machine learning. PMLR, 3020–3029

    2016

  5. [13]

    Ramesh Johari, Hannah Li, Inessa Liskovich, and Gabriel Y Weintraub. 2022. Experimental design in two-sided platforms: An analysis of bias.Management Science68, 10 (2022), 7069–7089

    2022

  6. [14]

    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

  7. [15]

    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

  8. [16]

    Christos Louizos, Uri Shalit, Joris M Mooij, David Sontag, Richard Zemel, and Max Welling. 2017. Causal effect inference with deep latent-variable models. Advances in neural information processing systems30 (2017)

    2017

  9. [17]

    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

  10. [18]

    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

  11. [19]

    Xinkun Nie and Stefan Wager. 2021. Quasi-oracle estimation of heterogeneous treatment effects.Biometrika108, 2 (2021), 299–319

    2021

  12. [20]

    2009.Causality

    Judea Pearl. 2009.Causality. Cambridge university press

    2009

  13. [21]

    Paul R Rosenbaum and Donald B Rubin. 1983. The central role of the propensity score in observational studies for causal effects.Biometrika70, 1 (1983), 41–55

    1983

  14. [22]

    Donald B Rubin. 1974. Estimating causal effects of treatments in randomized and nonrandomized studies.Journal of educational Psychology66, 5 (1974), 688

    1974

  15. [23]

    Piotr Rzepakowski and Szymon Jaroszewicz. 2010. Decision trees for uplift modeling. In2010 IEEE International Conference on Data Mining. IEEE, 441–450

    2010

  16. [24]

    Patrick Schwab, Lorenz Linhardt, and Walter Karlen. 2018. Perfect match: A simple method for learning representations for counterfactual inference with neural networks.arXiv preprint arXiv:1810.00656(2018)

    2018 arXiv

  17. [25]

    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

  18. [26]

    Claudia Shi, David Blei, and Victor Veitch. 2019. Adapting neural networks for the estimation of treatment effects.Advances in neural information processing systems32 (2019)

    2019

  19. [27]

    Zexu Sun and Xu Chen. 2024. M3TN: Multi-Gate Mixture-of-Experts Based Multi-Valued Treatment Network for Uplift Modeling. InICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). 5065–5069. doi:10.1109/ICASSP48485.2024.10446323

    2024 doi

  20. [28]

    Zexu Sun, Qiyu Han, Minqin Zhu, Hao Gong, Dugang Liu, and Chen Ma. 2025. Robust uplift modeling with large-scale contexts for real-time marketing. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 1. 1325–1336

    2025

  21. [29]

    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

  22. [30]

    Eric J Tchetgen Tchetgen and Tyler J VanderWeele. 2012. On causal inference in the presence of interference.Statistical methods in medical research21, 1 (2012), 55–75

    2012

  23. [31]

    Stefan Wager and Susan Athey. 2018. Estimation and inference of heterogeneous treatment effects using random forests.J. Amer. Statist. Assoc.113, 523 (2018), 1228–1242

    2018

  24. [32]

    Wenjie Wang, Fuli Feng, Xiangnan He, Liqiang Nie, and Tat-Seng Chua. 2021. Denoising implicit feedback for recommendation. InProceedings of the 14th ACM international conference on web search and data mining. 373–381

    2021

  25. [33]

    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. InProceedings of the 43rd International ACM SIGIR conference on research and develop...

    2020

  26. [34]

    Liuyi Yao, Sheng Li, Yaliang Li, Mengdi Huai, Jing Gao, and Aidong Zhang. 2018. Representation learning for treatment effect estimation from observational data. Advances in neural information processing systems31 (2018)

    2018

  27. [35]

    Jinsung Yoon, James Jordon, and Mihaela Van Der Schaar. 2018. GANITE: Esti- mation of individualized treatment effects using generative adversarial nets. In International conference on learning representations

    2018

  28. [36]

    Chenhao Zhai, Chang Meng, Xueliang Wang, Shuchang Liu, Xiaolong Hu, Shisong Tang, Xiaoqiang Feng, and Xiu Li. 2026. Heterogeneous Multi-treatment Uplift Modeling for Trade-off Optimization in Short-Video Recommendation. In Proceedings of the 32nd ACM SIGKDD Conference on Knowl...

    2026

  29. [37]

    Yan Zhao, Xiao Fang, and David Simchi-Levi. 2017. Uplift modeling with mul- tiple treatments and general response types. InProceedings of the 2017 SIAM International Conference on Data Mining. SIAM, 588–596

    2017

  30. [38]

    Zhe Zhao, Lichan Hong, Li Wei, Jilin Chen, Aniruddh Nath, Shawn Andrews, Aditee Kumthekar, Maheswaran Sathiamoorthy, Xinyang Yi, and Ed Chi. 2019. Recommending what video to watch next: a multitask ranking system. InPro- ceedings of the 13th ACM conference on recommender syste...

    2019

  31. [39]

    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. 46...

    2022

  32. [40]

    For each user𝑢, encode multi-source behaviors:𝐻 𝑢, ℎ𝑢 ←Encoder(𝑥 𝑢 )

  33. [41]

    For each candidate seller𝑠∈ S 𝑢 of user𝑢, compute candidate interaction:𝑧 𝑢,𝑠 ←TreatAttn(𝑠, 𝐻 𝑢 )

  34. [42]

    Predict control-side GMV:𝑝𝐺𝑀𝑉 𝑐 (𝑠) ←CtrlHead(𝑧 𝑢,𝑠 )

  35. [43]

    Predict platform-view GMV:𝑝𝐺𝑀𝑉 pla 𝑠 ←PlaHead(𝑧 𝑢,𝑠 )

  36. [44]

    Predict redemption probability:𝑝 𝑟 (𝑠) ←RedemHead(𝑧 𝑢,𝑠 )

  37. [45]

    Predict path-specific increments:𝑝Δ𝐺𝑀𝑉 𝑟 (𝑠), 𝑝Δ𝐺𝑀𝑉 1−𝑟 (𝑠) ←RDDHeads(𝑧 𝑢,𝑠 )

  38. [46]

    Combine the RDD prediction:𝑝𝐺𝑀𝑉 𝑠 ←𝑝𝐺𝑀𝑉 𝑐 (𝑠) +𝑝 𝑟 (𝑠)𝑝Δ𝐺𝑀𝑉 𝑟 (𝑠) + (1−𝑝 𝑟 (𝑠))𝑝Δ𝐺𝑀𝑉 1−𝑟 (𝑠)

  39. [47]

    Aggregate platform prediction:𝑝𝐺𝑀𝑉 pla 𝑢 ← Í 𝑠∈ S𝑢 𝑝𝐺𝑀𝑉 pla 𝑠

  40. [48]

    ComputeL seller,L pla, andL redem

  41. [49]

    user alternates between similar-category sellers

    Update parameters by minimizingL total. C.3 Why Single-Point Predictions Capture Global Effects The key insight is that the global aggregation constraint during trainingencourages the model to encode cross-seller substi- tution effectsinto the model parameters. Specifically: •...

Pith tools

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