Pith. sign in

REVIEW 4 major objections 7 minor 32 references

Fixed marketing budgets across channels should be reallocated by local causal slopes inside the data support, not by global predict-then-optimize.

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 ↗

T0 review · grok-4.5

2026-07-31 15:23 UTC pith:5KWPH7VP

load-bearing objection Solid industrial packaging of simplex reallocation with a real Taobao A/B; the online headline is a partial-deployment ITT and interference is the real caveat, not the theory. the 4 major comments →

arxiv 2607.28182 v1 pith:5KWPH7VP submitted 2026-07-30 cs.LG

Multi-channel Uplift Policy Learning

Changjian Liu , Tianyu Wang , Xiaoxuan Deng , WenTao Zhu , Yuwei Xu , Jungqi Jin , Yong Gao , Chuan Yu
show 2 more authors
Jian Xu Bo Zheng
This is my paper
classification cs.LG
keywords upliftresource allocationmulti-channel marketingcausal inferencepredict-then-optimizesimplex constraintssupport-aware policy
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.

E-commerce platforms must split a fixed marketing budget across channels such as ads and coupons. Standard practice first predicts outcomes then globally optimizes the predicted surface, but that fails when historical allocations are confounded and when the optimizer leaves the region where data actually exists. This paper reframes the problem as compositional uplift on the budget simplex: the useful signal is the relative marginal gain of moving money from one channel to another, not an absolute channel effect. ReAlloc recovers those local causal gradients with a fast orthogonal teacher on recent logs, distills them into a slow student potential whose differences score reallocations consistently, and then only accepts conservative local moves that stay inside empirically supported regions. Offline and online evidence on Taobao shows the approach raises both order volume and platform income while avoiding unsupported recommendations that pure optimizers produce.

Core claim

Under a fixed multi-channel budget, business value is the path integral of local causal reallocation gradients along feasible moves on the simplex. Learning and acting on a support-aware marginal field distilled from an orthogonal teacher yields higher deployable uplift than predict-then-optimize or independent-channel baselines, and in production simultaneously lifts pay orders and income.

What carries the argument

ReAlloc’s fast-slow loop: an orthogonal teacher residualizes outcomes and allocations to extract unbiased local simplex gradients; a student scalar potential distills those gradients into a path-consistent marginal field; support-aware local search only accepts reallocations whose conservative gain clears a safety threshold inside the logged support.

Load-bearing premise

Given the observed pre-decision state, the logged allocation is as good as random inside the supported region—if hidden factors drove both past budgets and outcomes, the teacher’s local gradients stay biased.

What would settle it

A randomized multi-channel budget experiment where ReAlloc’s recommended local moves, restricted to the same support rule, fail to beat the incumbent logging policy on doubly robust uplift, or produce higher support-violation rates and lower deployable uplift than support-constrained S-learner baselines under the paper’s own hard-overlap synthetic regimes.

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

If this is right

  • Global black-box optimization of factual response models is unsafe for fixed-budget multi-channel marketing; deployable value comes from local, support-constrained reallocation.
  • Cross-channel substitution can be captured by a distilled marginal field without reconstructing a full global response surface.
  • A fast teacher on short windows plus a slow student memory can retain local geometry under rotating support at far lower cost than pooled full-history retraining.
  • Production systems can raise order volume and profitability together by intervening less often but only on empirically supported moves.

Where Pith is reading between the lines

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

  • The same simplex-local-gradient plus support gate pattern likely transfers to other zero-sum compositional decisions (inventory slots, compute quotas, multi-touch creative budgets) where global PTO over-extrapolates.
  • If unobserved confounding is material, pairing ReAlloc’s support layer with explicit exploration or instrumental variation would be the natural next stress test the paper’s offline DR setup already partially enables.
  • The observed GMV-versus-orders trade-off suggests the learned field may be optimizing conversion intensity more than basket size; multi-objective potentials could make that trade-off explicit.

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

4 major / 7 minor

Summary. The paper formulates fixed-budget multi-channel marketing allocation as a simplex-constrained compositional uplift problem, where the decision primitive is the relative marginal return of zero-sum reallocations rather than absolute ITEs. It proposes ReAlloc, a fast–slow framework: an orthogonal teacher recovers local causal gradients from short-term logs (Robinson-style residualization), a potential-parameterized student distills those gradients into a path-consistent marginal field, and a support-aware local search executes conservative updates. Theory links path integrals of the causal reallocation field to uplift, gives Neyman-orthogonal local-field rates, and bounds regret by field error times path length. Synthetic DGPs stress confounding, overlap, and extrapolation; offline Taobao evaluations use matched residual ranking and DR OPE; a 14-day item-level A/B test (300K items, 10% treated) reports +3.53% pay orders and +3.26pt platform income versus AdditiveROI, with lower spend and a GMV decline.

Significance. If the results hold under realistic marketplace interference, the work is a substantial contribution to industrial causal decision-making: it correctly reframes multi-channel budgeting as compositional uplift on the simplex, diagnoses concrete PTO failure modes (confounding, objective mismatch, extrapolation), and ships a deployable teacher–student design with support conservatism. Strengths include coherent theory (Thm. 5.1 path-integral uplift; Thm. 5.2 orthogonal local rates; Thm. 5.3 regret via potential accuracy), a synthetic DGP that isolates confounding/overlap/OOS with ablations (Tables 1–3), and a large-scale online A/B with simultaneous order and income lifts (Table 5). The support-aware decision layer is well motivated and empirically decisive offline (Table 4: support violations drop from ~0.68 to 0). These are genuine engineering and methodological advances for constrained multi-treatment uplift.

major comments (4)
  1. [§7.2, Table 5] §7.2 and Table 5: The headline claim rests on an item-level A/B versus AdditiveROI (+3.53% pay orders, +3.26pt income). The manuscript asserts that “strict item-level randomization prevents budget interference,” which blocks shared per-item budget-pool leakage but not the spillovers that matter for multi-channel marketing: paid-ad reallocations enter shared auctions (bids/impressions of treated items affect controls), and benefit/coupon mix shifts alter cross-item substitution in user choice. With only 10% treated, Table 5 identifies a partial-deployment ITT, not necessarily the full-rollout equilibrium effect. Please either (i) provide interference diagnostics (e.g., auction-level exposure, neighbor/category spillover tests, dose–response by local treatment density) or (ii) explicitly reframe the online claim as a partial-deployment effect and discuss sign/magnitude risk under full roll
  2. [Table 4B, §7.1–7.2] Table 4B: On randomized exploration traffic, ReAlloc’s DR lift (.025) is nearly identical to AdditiveROI (.023), VCNet (.022), and GIKS (.022); the main offline separation is support violation rate and per-acted lift under lower action rate (.33 vs ~.97). The online superiority over AdditiveROI is therefore only weakly prefigured by aggregate causal value offline. Please strengthen mechanism attribution: report whether online gains concentrate on the support-constrained, low-action regime predicted by Table 4; add channel-mix / reallocation-path diagnostics; and clarify how much of the A/B lift is attributable to support conservatism versus better local geometry. Without this, the causal story linking ReAlloc’s design to Table 5 remains thin.
  3. [§4.1, §5.2, Assumption 1 / A.4] Assumption 1 (App. A.1) and §5.2/A.4: The teacher is described as extracting “unbiased local gradients” (§4.1, abstract). The paper correctly notes that orthogonalization removes only the observed propensity component and does not recover omitted confounders (shortcut bias b_sc). In production logs, legacy policies almost surely depend on unobserved item/market factors that also drive Y. Please temper causal language around the teacher throughout (abstract, §4.1, contributions) to “orthogonalized / debiased w.r.t. observed assignment,” and state clearly which claims (offline DR, matched ρ) require ignorability and which (A/B contrast) do not. This is not fatal to the online experiment but is load-bearing for interpreting Stages I–II as causal.
  4. [Table 5, §7.2, §8] Table 5: Pay orders rise while GMV falls (−2.6%) and total cost falls (−2.5%), with ROI statistically flat. The conclusion notes a “GMV trade-off” but does not analyze whether this is intended (more low-AOV conversions), a horizon effect, or a channel-mix artifact (e.g., coupons vs ads). For a platform utility paper whose abstract claims simultaneous lifts in “pay order and income,” please report AOV, channel spend shares, and income definition, and discuss whether the objective in (2) matches the business utility that includes GMV. Otherwise the multi-objective success claim is incomplete.
minor comments (7)
  1. [Abstract, §4.2] Abstract and §4 title the student “Explanation-Guided”; the body describes finite-difference / Jacobian distillation into a potential. Align terminology (explanation-guided vs marginal-field distillation) for consistency.
  2. [Figure 3] Figure 3 caption: “Orcle” → “Oracle”.
  3. [§4.1, Eq. (11)] Eq. (11) vs (9)–(10): clarify whether the teacher gradient regularizer is evaluated only at z=0 or along residualized ˜p; the text suggests local directional derivative at the anchor while the loss uses ⟨g_ψ, ˜p⟩.
  4. [§5.2] Proposition 1: “qality” → “quality” in the title.
  5. [Algorithm 1, §4] Several free hyperparameters (λ_res, λ_grad, λ_pair, λ_jac, β, λ_s, τ_min, γ, B, h) are listed in Algorithm 1 / method text without sensitivity ranges used in the Taobao deployment. A short appendix table would aid reproducibility.
  6. [§2] Related work on continuous/multi-treatment HTE and budgeted uplift is adequate; a brief pointer to offline RL conservative objectives beyond CQL (already cited) and to compositional data analysis (ILR) would help readers map the simplex geometry.
  7. [Table 1] Table 1 superscript notation for raw uplift on unconstrained PTO is easy to miss; consider a separate Raw column or clearer footnote.

Circularity Check

0 steps flagged

No significant circularity: theory is path-integral identities under stated assumptions; empirics are external A/B and oracle evaluations, not fitted-then-predicted quantities.

full rationale

The paper’s load-bearing chain does not collapse inputs into claimed outputs by construction. Theorem 5.1 equates simplex uplift to the path integral of the causal reallocation field under continuous differentiability—an FTC-style identity, not a fit. Theorems 5.2–5.3 bound orthogonal local-field error and regret under explicit identification, overlap, smoothness, and approximation assumptions (A.1–A.5); the bounds are not tautological restatements of the training loss. The teacher–student pipeline is supervised distillation of finite differences/Jacobians into a scalar potential (Eqs. 14–15, 12–13), presented as transfer for stability, not as an independent first-principles derivation of policy value. Support-aware search and the deployable-uplift metric with OOS fallback are design/safety choices evaluated against synthetic oracles and Taobao A/B (Table 5) and DR/matched ranks (Table 4)—external contrasts, not quantities forced by the fitting objective. Related-work self-citations (e.g., marketing hosting) are contextual, not uniqueness theorems that forbid alternatives or force the main claim. No self-definitional loop, fitted-input-as-prediction, or renaming of a known result as a derived law was found.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 2 invented entities

Load-bearing content is standard causal identification plus engineering choices for local estimation and conservative search. The online claim rests on randomized item assignment and business outcome definitions; the method claim rests on ignorability/overlap/smoothness and on hyperparameters that gate which moves are allowed.

free parameters (6)
  • Teacher loss weights λ_res, λ_grad
    Balance factual fit, residual fit, and directional derivative regularization in L_teacher (Eq. 11); not derived, chosen for training stability.
  • Student loss weights λ_pair, λ_jac
    Trade pairwise finite-difference distillation vs Jacobian matching in L_student (Eq. 15).
  • Conservative decision penalties β, λs and threshold τ_min
    Scale uncertainty and support penalties in G_cons and decide whether to no-op (Eqs. 18–19); directly control blast radius and deployable uplift.
  • Student EMA rate γ and replay buffer size B = B≈n in reported config
    Control fast-slow memory and compute/storage trade-off in temporal experiments (§4.2, RQ3).
  • Local step set S / δ candidates and R_max
    Discretize feasible simplex transfers and cap multi-step search depth; affects regret term ε_search.
  • Kernel bandwidth h / support density model
    Local orthogonal estimation rate and path-support gate depend on bandwidth and nonconformity threshold ε_S (Thm 5.2, App. B.5).
axioms (5)
  • domain assumption Conditional ignorability / identification on supported actions: {Y(p)} ⊥ P | X for p in C(X), and Y=Y(P).
    Assumption 1 (App. A.1); required for teacher gradients to equal causal reallocation field g★ rather than observational shortcut field g_obs.
  • domain assumption Local overlap: kernel mass around supported (X,z) scales as Θ(h^d) uniformly on R.
    Assumption 2; underpins finite-sample rate ε_T in Theorem 5.2 and deployability of local search.
  • domain assumption Response continuously differentiable on C(X) with controlled local remainder so path integrals of g★ recover Δ.
    Assumption 3 + Theorem 5.1; justifies potential-student parameterization and regret bound 2 L_max ε_g.
  • standard math Cross-fitting nuisance estimation with empirical process bound on the orthogonal score.
    Assumption 4; standard DML-style condition used for Neyman orthogonality and rate decomposition.
  • ad hoc to paper Implemented teacher/student stay within ε_comp and ε_S of the orthogonal population field on R.
    Assumption 5 bridges neural composite objectives to the population orthogonal root; not independently verified beyond ablations.
invented entities (2)
  • Causal reallocation field g★ on the simplex tangent space independent evidence
    purpose: Replace absolute channel ITEs with relative marginal returns under zero-sum budget moves.
    Defined as Π_T ∇_p μ; standard differential construction on the simplex rather than a new physical object, but it is the paper’s central decision primitive.
  • Fast-slow ReAlloc teacher–student marginal field with support-aware conservative gain independent evidence
    purpose: Convert short-window orthogonal local slopes into a stable long-horizon policy that refuses unsupported extrapolation.
    Methodological system (Stages I–III); validated via synthetic oracle metrics and Taobao A/B, not an extra latent substance.

pith-pipeline@v1.2.0-daily-grok45 · 26878 in / 3874 out tokens · 84914 ms · 2026-07-31T15:23:59.297832+00:00 · methodology

0 comments
read the original abstract

E-commerce platforms must allocate fixed marketing budgets across multiple channels to maximize business utility. However, standard predict-then-optimize (PTO) paradigms fail in this compositional space due to observational confounding and severe extrapolation. We formulate this challenge as a simplex-constrained uplift decision problem and propose ReAlloc, a fast-slow causal framework. Specifically, an agile Orthogonal Teacher extracts unbiased local gradients from short-term logs, while an Explanation-Guided Student distills them into a structured marginal field over long-term horizons. This design enables support-aware, conservative decisions that capture cross-channel substitutions. Extensive simulations and large-scale online A/B tests on Taobao platform demonstrate that ReAlloc achieves simultaneous lifts in both pay order and income.

Figures

Figures reproduced from arXiv: 2607.28182 by Bo Zheng, Changjian Liu, Chuan Yu, Jian Xu, Jungqi Jin, Tianyu Wang, Wentao Zhu, Xiaoxuan Deng, Yong Gao, Yuwei Xu.

Figure 1
Figure 1. Figure 1: Conceptual illustration of ReAlloc. (a) Local response slopes identify promising reallocation directions. (b) Global PTO extrapolation produces unreliable decisions. (c) ReAlloc composes local slopes within support, leading to better decisions. To address these challenges, we propose ReAlloc, a framework targeting local reallocation for multi-channel uplift decision prob￾lem. It learns the causal marginal … view at source ↗
Figure 2
Figure 2. Figure 2: Overview of ReAlloc 4.1 Stage I: Orthogonal Response Teacher To address observational confounding (Challenge 1), we use orthog￾onalization, akin to Robinson’s transformation in double machine learning (DML). For each recent training window Dshort, we first estimate two nuisance functions via cross-fitting: 𝑚ˆ (𝐻, 𝐵) ≈ E[𝑌 | 𝐻, 𝐵], 𝑒ˆ(𝐻, 𝐵) ≈ E[𝑝 | 𝐻, 𝐵], (7) where 𝑚ˆ captures the baseline demand and 𝑒ˆ ∈ Δ… view at source ↗
Figure 3
Figure 3. Figure 3: Orcle response surface and policy behavior. [PITH_FULL_IMAGE:figures/full_fig_p007_3.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

32 extracted references · 3 canonical work pages

  1. [1]

    Meng Ai, Biao Li, Heyang Gong, Qingwei Yu, Shengjie Xue, Yuan Zhang, Yunzhou Zhang, and Peng Jiang. 2022. LBCF: A Large-Scale Budget-Constrained Causal Forest Algorithm. InProceedings of the ACM Web Conference 2022(Virtual Event, Lyon, France)(WWW ’22). Association for Computing Machinery, New York, NY, USA, 2310–2319. doi:10.1145/3485447.3512103

    arXiv 2022

  2. [2]

    Javier Albert and Dmitri Goldenberg. 2022. E-Commerce Promotions Personaliza- tion via Online Multiple-Choice Knapsack with Uplift Modeling. InProceedings of the 31st ACM International Conference on Information & Knowledge Management (Atlanta, GA, USA)(CIKM ’22). Association for Computing Machinery, New York, NY, USA, 2863–2872. doi:10.1145/3511808.3557100

    arXiv 2022

  3. [3]

    Susan Athey and Stefan Wager. 2021. Policy Learning with Observational Data. Econometrica89, 1 (2021), 133–161. doi:10.3982/ECTA15732 Multi-channel Uplift Policy Learning

    doi:10.3982/ecta15732 2021

  4. [4]

    Victor Chernozhukov, Denis Chetverikov, Mert Demirer, Esther Duflo, Christian Hansen, Whitney Newey, and James Robins. 2018. Double/Debiased Machine Learning for Treatment and Structural Parameters.The Econometrics Journal21, 1 (2018), C1–C68. doi:10.1111/ectj.12097

    doi:10.1111/ectj.12097 2018

  5. [5]

    Yuan Deng, Negin Golrezaei, Patrick Jaillet, Jason Cheuk Nam Liang, and Vahab Mirrokni. 2023. Multi-channel Autobidding with Budget and ROI Constraints. InProceedings of the 40th International Conference on Machine Learning (Proceed- ings of Machine Learning Research, Vol. 202), Andreas Krause, Emma Brunskill, Kyunghyun Cho, Barbara Engelhardt, Sivan Saba...

    2023

  6. [6]

    Floris Devriendt, Jente Van Belle, Tias Guns, and Wouter Verbeke. 2022. Learn- ing to Rank for Uplift Modeling.IEEE Transactions on Knowledge and Data Engineering34, 10 (2022), 4888–4904. doi:10.1109/TKDE.2020.3048510

    arXiv 2022

  7. [7]

    Miroslav Dudík, Dumitru Erhan, John Langford, and Lihong Li. 2014. Doubly Robust Policy Evaluation and Optimization.Statist. Sci.29, 4 (2014), 485–511. doi:10.1214/14-STS500

    doi:10.1214/14-sts500 2014

  8. [8]

    Miroslav Dudík, John Langford, and Lihong Li. 2011. Doubly robust policy evalu- ation and learning. InProceedings of the 28th International Conference on Interna- tional Conference on Machine Learning(Bellevue, Washington, USA)(ICML’11). Omnipress, Madison, WI, USA, 1097–1104

    2011

  9. [9]

    Predict, then Optimize

    Adam N. Elmachtoub and Paul Grigas. 2022. Smart “Predict, then Optimize”. Management Science68, 1 (2022), 9–26. doi:10.1287/mnsc.2020.3922

    arXiv 2022

  10. [10]

    Sahin Cem Geyik, Abhishek Saxena, and Ali Dasdan. 2015. Multi-Touch Attribution Based Budget Allocation in Online Advertising.arXiv preprint arXiv:1502.06657(2015)

    Pith/arXiv arXiv 2015

  11. [11]

    Pierre Gutierrez and Jean-Yves Gérardy. 2017. Causal Inference and Uplift Modelling: A Review of the Literature. InProceedings of The 3rd International Conference on Predictive Applications and APIs (Proceedings of Machine Learning Research, Vol. 67), Claire Hardgrove, Louis Dorard, Keiran Thompson, and Florian Douetteau (Eds.). PMLR, 1–13

    2017

  12. [12]

    Keisuke Hirano and Guido W. Imbens. 2004.The Propensity Score with Continuous Treatments. John Wiley & Sons, Ltd, Chapter 7, 73–84. doi:10.1002/0470090456. ch7

    work page doi:10.1002/0470090456 2004

  13. [13]

    Kennedy, Zongming Ma, Matthew D

    Edward H. Kennedy, Zongming Ma, Matthew D. McHugh, and Dylan S. Small

  14. [14]

    Aviral Kumar, Aurick Zhou, George Tucker, and Sergey Levine. 2020. Conservative Q-Learning for Offline Reinforcement Learning. InAdvances in Neural Information Processing Systems, H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin (Eds.), Vol. 33. Curran Associates, Inc., 1179–1191. https://proceedings.neurips.cc/paper_files/paper/2020/file/ 0...

    2020

  15. [15]

    Sachin Kumar, Garima Gupta, Ranjitha Prasad, Arnab Chatterjee, Lovekesh Vig, and Gautam Shroff. 2020. CAMTA: Causal Attention Model for Multi-touch Attribution.arXiv preprint arXiv:2012.11403(2020)

    Pith/arXiv arXiv 2020

  16. [16]

    Künzel, Jasjeet S

    Sören R. Künzel, Jasjeet S. Sekhon, Peter J. Bickel, and Bin Yu. 2019. Met- alearners for Estimating Heterogeneous Treatment Effects Using Machine Learn- ing.Proceedings of the National Academy of Sciences116, 10 (2019), 4156–4165. doi:10.1073/pnas.1804597116

    doi:10.1073/pnas.1804597116 2019

  17. [17]

    Xinkun Nie and Stefan Wager. 2021. Quasi-Oracle Estimation of Heterogeneous Treatment Effects.Biometrika108, 2 (2021), 299–319. doi:10.1093/biomet/asaa076

    doi:10.1093/biomet/asaa076 2021

  18. [18]

    Diego Olaya, Kristof Coussement, and Wouter Verbeke. 2020. A Survey and Benchmarking Study of Multitreatment Uplift Modeling.Data Mining and Knowledge Discovery34, 2 (2020), 273–308. doi:10.1007/s10618-019-00670-y

    doi:10.1007/s10618-019-00670-y 2020

  19. [19]

    Nicholas Radcliffe. 2007. Using Control Groups to Target on Predicted Lift: Building and Assessing Uplift Model.Direct Marketing Analytics Journal(2007), 14–21

    2007

  20. [20]

    Utsav Sadana, Abhilash Chenreddy, Erick Delage, Alexandre Forel, Emma Fre- jinger, and Thibaut Vidal. 2025. A Survey of Contextual Optimization Methods for Decision-Making under Uncertainty.European Journal of Operational Research 320, 2 (2025), 271–289. doi:10.1016/j.ejor.2024.03.020

    doi:10.1016/j.ejor.2024.03.020 2025

  21. [21]

    Johansson, and David Sontag

    Uri Shalit, Fredrik D. Johansson, and David Sontag. 2017. Estimating individual treatment effect: generalization bounds and algorithms. InProceedings of the 34th International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 70), Doina Precup and Yee Whye Teh (Eds.). PMLR, 3076–3085. https://proceedings.mlr.press/v70/shalit17a.html

    2017

  22. [22]

    Guangyuan Shen, Shenjie Sun, Dehong Gao, Shaolei Li, Libin Yang, Yongping Shi, and Wei Ning. 2023. Cross-channel Budget Coordination for Online Advertising System.arXiv preprint arXiv:2305.06883(2023)

    Pith/arXiv arXiv 2023

  23. [23]

    Claudia Shi, David Blei, and Victor Veitch. 2019. Adapting Neural Networks for the Estimation of Treatment Effects. InAdvances in Neural Information Processing Systems, H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett (Eds.), Vol. 32. Curran Associates, Inc. https://proceedings.neurips. cc/paper_files/paper/2019/file/8fb5f...

    2019

  24. [24]

    Zexu Sun, Hao Yang, Dugang Liu, Yunpeng Weng, Xing Tang, and Xiuqiang He. 2024. End-to-End Cost-Effective Incentive Recommendation under Budget Constraint with Uplift Modeling. InProceedings of the 18th ACM Conference on Recommender Systems. 560–569. doi:10.1145/3640457.3688147

    arXiv 2024

  25. [25]

    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. doi:10.1080/01621459.2017.1319839

    arXiv 2018

  26. [26]

    Bingzhe Wang, Tianyu Wang, Qi Qi, Xiaoxuan Deng, Zhilin Zhang, and Chuan Yu. 2026. Marketing Hosting: From Fixed to Endogenous Budgets. InProceedings of the ACM Web Conference 2026(United Arab Emirates)(WWW ’26). Association for Computing Machinery, New York, NY, USA, 327–338. doi:10.1145/3774904. 3792519

    doi:10.1145/3774904 2026

  27. [27]

    Bryan Wilder, Bistra Dilkina, and Milind Tambe. 2019. Melding the Data- Decisions Pipeline: Decision-Focused Learning for Combinatorial Optimization. Proceedings of the AAAI Conference on Artificial Intelligence33, 01 (Jul. 2019), 1658–1665. doi:10.1609/aaai.v33i01.33011658

    doi:10.1609/aaai.v33i01.33011658 2019

  28. [28]

    Williams and Catherine M

    Justin R. Williams and Catherine M. Crespi. 2020. Causal Inference for Multiple Continuous Exposures via the Multivariate Generalized Propensity Score.arXiv preprint arXiv:2008.13767(2020). doi:10.48550/arXiv.2008.13767

    Pith/arXiv arXiv doi:10.48550/arxiv.2008.13767 2020

  29. [29]

    Yan Zhao, Xiao Fang, and David Simchi-Levi. 2017. Uplift Modeling with Multiple Treatments and General Response Types. InProceedings of the 2017 SIAM Interna- tional Conference on Data Mining. SIAM, 588–596. doi:10.1137/1.9781611974973.66

    work page doi:10.1137/1.9781611974973.66 2017

  30. [30]

    Zhenyu Zhao and Totte Harinen. 2019. Uplift Modeling for Multiple Treatments with Cost Optimization. In2019 IEEE International Conference on Data Science and Advanced Analytics (DSAA). IEEE, 422–431. doi:10.1109/DSAA.2019.00057

    arXiv 2019

  31. [31]

    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(Washington DC, USA) (KDD ’22). Association for Computing Machinery, New York, NY, U...

    arXiv 2022

  32. [2017]

    doi:10.1111/rssb.12212

    Non-parametric Methods for Doubly Robust Estimation of Continuous Treatment Effects.Journal of the Royal Statistical Society: Series B79, 4 (2017), 1229–1245. doi:10.1111/rssb.12212

    work page doi:10.1111/rssb.12212 2017