Paper deep dive
From Prediction to Incrementality: Causal Optimization for Large-Scale Targeting and Recommendation
Changshuai Wei, John Bencina, Phuc Nguyen, Andre Assuncao Silva T Ribeiro, Benjamin Zelditch
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 91%
Last extracted: 8/13/2026, 4:39:39 AM
Summary
The paper presents a decision-centric framework for large-scale targeting and recommendation that optimizes causal effects under global constraints, addressing the misallocation of resources inherent in predictive-only systems. The framework integrates a causal neural network with a Transformer backbone for individual treatment-effect estimation, a Bayesian neural-bandit layer for uncertainty-aware exploration, and a dual-based large-scale linear-programming layer for constrained allocation. Evaluated on LinkedIn Feed marketing traffic, the system achieved a statistically significant 7.20% lift in long-term value.
Entities (8)
Relation Signals (6)
Framework → achievedlift → 7.20%
confidence 95% · The end-to-end treatment policy delivered a statistically significant +7.20% lift in the primary long-term-value metric
Framework → evaluatedon → LinkedIn
confidence 95% · online A/B test on LinkedIn Feed marketing traffic
DragonNet → extendedby → Transformer
confidence 90% · We extend DragonNet with three key enhancements... First, we feed sequential user-event features through a transformer layer
Bayesian Neural-Bandit → providesinputto → Linear Programming
confidence 90% · Using the sampled incremental objective tau~ from DragonBandit, the horizon objective is max sum tau~ x
Linear Programming → solvedby → Dual Decomposition
confidence 90% · we solve the resulting large-scale problem using a smoothed dual-decomposition method
Bayesian Neural-Bandit → usesalgorithm → Thompson Sampling
confidence 90% · Thompson sampling proceeds by drawing a logit sample... DragonBandit draws posterior incremental scores
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Large-scale targeting and recommendation systems are typically built around predictive scores fed into heuristic or local allocation. When the business goal is incremental impact, as in marketing campaigns, incentives, and notifications, this paradigm systematically misallocates resources toward users who would have acted anyway. We present a decision-centric framework that instead optimizes causal effects under global constraints, aligning three components under a single objective: a causal neural network with a Transformer backbone for individual treatment-effect estimation, a Bayesian neural-bandit layer for uncertainty-aware exploration, and a dual-based large-scale linear-programming layer for constrained allocation. The framework also supports sequential context and multi-outcome, attribute-conditioned scoring through a Transformer encoder and outcome embeddings. We evaluate it with offline simulations on a public bandit dataset, targeted architectural ablations, and an online A/B test on LinkedIn Feed marketing traffic. We also distill production lessons on causal training-data construction and cost and delivery control, which were critical to successful deployment. The end-to-end treatment policy delivered a statistically significant $+7.20\%$ lift in the primary long-term-value metric, demonstrating the feasibility of production-scale causal optimization under business constraints.
Tags
Links
- Source: https://arxiv.org/abs/2608.10182v1
- Canonical: https://arxiv.org/abs/2608.10182v1
Trouble viewing inline? Open PDF directly →
Full Text
60,167 characters extracted from source content.
Expand or collapse full text
From Prediction to Incrementality: Causal Optimization for Large-Scale Targeting and Recommendation Changshuai Wei chawei@linkedin.com LinkedInSeattleUSA , John Bencina jbencina@linkedin.com LinkedInSunnyvaleUSA , Phuc Nguyen honnguyen@linkedin.com LinkedInSunnyvaleUSA , Andre Assuncao Silva T Ribeiro aribeiro@linkedin.com LinkedInSunnyvaleUSA and Benjamin Zelditch bzelditch@linkedin.com LinkedInSunnyvaleUSA (2026) Abstract. Large-scale targeting and recommendation systems are typically built around predictive scores fed into heuristic or local allocation. When the business goal is incremental impact, as in marketing campaigns, incentives, and notifications, this paradigm systematically misallocates resources toward users who would have acted anyway. We present a decision-centric framework that instead optimizes causal effects under global constraints, aligning three components under a single objective: a causal neural network with a Transformer backbone for individual treatment-effect estimation, a Bayesian neural-bandit layer for uncertainty-aware exploration, and a dual-based large-scale linear-programming layer for constrained allocation. The framework also supports sequential context and multi-outcome, attribute-conditioned scoring through a Transformer encoder and outcome embeddings. We evaluate it with offline simulations on a public bandit dataset, targeted architectural ablations, and an online A/B test on LinkedIn Feed marketing traffic. We also distill production lessons on causal training-data construction and cost and delivery control, which were critical to successful deployment. The end-to-end treatment policy delivered a statistically significant +7.20%+7.20\% lift in the primary long-term-value metric, demonstrating the feasibility of production-scale causal optimization under business constraints. Causal ML, Transformer, Bandit, Linear Program †copyright: acmlicensed†journalyear: 2026†doi: X.X†conference: X; X; X†isbn: X†ccs: Mathematics of computing Probability and statistics†ccs: Computing methodologies Machine learning†ccs: Information systems Information retrieval 1. Introduction Large-scale targeting and recommendation systems are a core component of modern online platforms, supporting applications such as advertising, marketing outreach, notifications, and content recommendation. In practice, these systems are typically built by training predictive models to estimate user response probabilities (such as click-through or conversion likelihood) and then ranking or selecting items based on these predictions. This paradigm has proven effective for optimizing engagement metrics at scale and has become the dominant approach in both industrial deployments and academic research. However, predictive modeling alone is fundamentally misaligned with the objectives of many real-world targeting problems. In marketing and incentive-driven settings, the goal is not to predict outcomes under historical policies, but to estimate the incremental impact of an intervention relative to a counterfactual baseline. Observed user responses are confounded by prior targeting decisions, exposure mechanisms, and user self-selection, implying that high predicted response does not necessarily correspond to high causal value. Prior work has demonstrated that learning and evaluation based on biased recommendation logs can systematically misestimate the true effect of interventions when deployed for decision-making (Schnabel et al., 2016; Wang et al., 2018). This challenge has motivated increasing interest in causal machine learning for recommender and targeting systems. A line of work reframes recommendations as treatments and user interactions as potential outcomes, enabling principled correction for exposure bias and confounding (Schnabel et al., 2016; Bonner and Vasile, 2018). More broadly, causal perspectives on recommendation have been surveyed extensively, highlighting the limitations of purely predictive approaches and emphasizing the role of causal inference in reliable decision-making (Gao et al., 2024). In marketing and targeting applications, uplift modeling and treatment effect estimation explicitly target incrementality by estimating individual-level causal effects rather than response probabilities. Foundational work has studied the theoretical properties of individual treatment effect estimation and proposed representation-learning-based approaches to mitigate confounding bias (Shalit et al., 2017). More recent neural architectures, such as DragonNet, jointly model potential outcomes and treatment assignment, demonstrating improved stability and accuracy in treatment effect estimation (Shi et al., 2019). Complementary to causal modeling, a rich literature studies counterfactual evaluation and learning from logged bandit feedback. Off-policy evaluation methods enable unbiased estimation of policy value without exhaustive online experimentation and have been extended to large action spaces and slate recommendation settings (Swaminathan and Joachims, 2017; McInerney et al., 2020). While these approaches provide principled tools for evaluation and exploration, they typically focus on estimating or comparing policies rather than optimizing decisions for incremental outcomes. At the same time, real-world targeting and recommendation systems must operate under complex global constraints, including budgets, capacity limits, frequency caps, and coverage or fairness requirements. To address these challenges, many industrial platforms rely on large-scale constrained optimization, often formulated as linear or mixed-integer programs, to translate model outputs into coordinated decisions. Theoretical foundations for robust and constrained optimization are well established (Bertsimas et al., 2011), and industrial systems have demonstrated the effectiveness of combining learned utility models with constraint-aware optimization to improve system-level objectives (Agarwal et al., 2015; Makhijani et al., 2019; Wei et al., 2024). Despite their success, existing optimization-based systems almost universally rely on predictive scores or locally estimated uplift values as inputs. When incrementality matters, optimizing constrained decisions using non-causal or locally optimal signals can systematically misallocate resources toward users who would have acted regardless of intervention, thereby reducing true return on investment. Conversely, causal effect estimates alone are insufficient without a principled optimization layer to enforce constraints and optimize system-level objectives. We present a decision-centric formulation in which incrementality and constraints are addressed jointly: the system optimizes causal effects under global constraints, replacing the more common pattern of optimizing predictive or proxy signals under heuristic allocation. This reframes targeting as choosing users with high incremental impact and as globally coordinated allocation, rather than as ranking high-response users. The three components fit together naturally under this objective: causal modeling defines what is optimized, uncertainty-aware exploration shapes the data on which the causal estimates are learned, and constrained optimization translates those estimates into feasible decisions under shared resources. We instantiate the framework with a Transformer-augmented DragonNet causal head, a Bayesian neural-bandit layer, and a large-scale LP layer. While our primary motivation arises from marketing, the formulation is not domain-specific and extends to other settings that allocate limited intervention capacity under constraints. Our contributions are: (i) a decision-centric formulation that aligns causal estimation, exploration, and constrained allocation under a single objective; (i) two DragonNet architectural extensions, a Transformer encoder for marketing-touchpoint sequences and a shared outcome-embedding head, for multi-outcome and attribute-conditioned incremental scoring; and (i) a productionized pipeline combining the causal head with neural-bandit exploration and a dual-based large-scale LP solver, evaluated on a public bandit dataset and as an end-to-end policy in a large online A/B test on LinkedIn Feed marketing traffic that delivered a +7.20%+7.20\% lift in the primary KPI (p=0.041p=0.041). 1.1. Related Work Recent work has begun to incorporate treatment effect estimation and uplift modeling directly into recommender systems, moving beyond purely predictive objectives. Chen et al. (Chen et al., 2024) propose using individual treatment effect estimation to guide user interest exploration, demonstrating that causal signals can improve exploration efficiency compared to prediction-based uncertainty. Meng et al. (Meng et al., 2025) introduce a coarse-to-fine dynamic uplift modeling framework for real-time video recommendation, showing that explicit uplift estimation can improve ranking quality at scale. Sun et al. (Sun et al., 2024) study incentive recommendation under budget constraints and propose an end-to-end uplift-based framework to improve cost-effectiveness. While these approaches demonstrate the practical value of causal and uplift modeling in recommender systems, they primarily focus on local decision-making, such as improving ranking quality or applying uplift-based heuristics under relatively simple constraints. Exploration is typically handled implicitly or through heuristic mechanisms, and optimization is performed at the level of individual users or items rather than as a coordinated system-level problem. In contrast, our work integrates causal effect estimation, principled exploration, and constrained optimization into a unified decision framework. We combine DragonNet-style causal modeling with Transformer-based sequential representations and a neural bandit layer, enabling uncertainty-aware exploration guided by estimated incremental effects. On the optimization side, we formulate targeting decisions as a large-scale constrained optimization problem and solve it using a dual-based approach that scales to extreme problem sizes. This allows us to coordinate decisions across users, items, and campaigns while enforcing complex global constraints, going beyond ranking-based or locally optimal uplift methods. 2. Incremental Optimization Framework 2.1. Incremental Modeling Figure 1. Incremental model architecture overview. The model extends DragonNet with a transformer for temporal modeling and outcome embeddings for multi-product prediction. The core model utilizes member-level embeddings, dense features, and the marketing interaction sequence. The auxiliary autoencoder model utilizes product-level embeddings corresponding to each modeled outcome. Incremental optimization requires estimating user-level treatment effects that quantify the expected lift from a targeting or recommendation action. Under the potential outcomes framework, let Y(1)Y(1) and Y(0)Y(0) denote the potential outcomes under treatment (T=1T=1) and control (T=0T=0), and let X∈X denote observed covariates. The Conditional Average Treatment Effect (CATE), or Individual Treatment Effect (ITE) at the covariate level, is (1) τ(X)=[Y(1)−Y(0)∣X]=μ1(X)−μ0(X),τ(X)\;=\;E[Y(1)-Y(0) X]\;=\; _1(X)- _0(X), where μt(X)=[Y(t)∣X] _t(X)=E[Y(t) X]. Two standard assumptions identify τ(X)τ(X) from observational data: unconfoundedness, Y(0),Y(1)⟂T∣X\Y(0),Y(1)\ \!\!\! T X, and overlap, 0<e(X)<10<e(X)<1 for all X, where e(X)=P(T=1∣X)e(X)=P(T=1 X) is the propensity score. Under these, μt(X)=[Y∣X,T=t] _t(X)=E[Y X,T=t] and τ(X)τ(X) becomes point-identified from the observed data distribution. Various estimators for τ(X)τ(X) have been proposed, including S-Learner, T-Learner, X-Learner, and Double Machine Learning (DML) (Künzel et al., 2019; Chernozhukov et al., 2018). These approaches typically require multi-stage modeling pipelines, which introduce operational complexity at our scale of hundreds of millions of members. We instead adopt a modified version of DragonNet (Shi et al., 2019), shown in Figure 1, which jointly estimates μ1(X) _1(X), μ0(X) _0(X), and e(X)e(X) in a single forward pass. DragonNet extends TARNET (Shalit et al., 2017) by adding a propensity head to a shared-representation architecture: a shared encoder Φ feeds two outcome heads estimating μ1 _1 and μ0 _0 (with observed-treatment gradient routing), and a third head estimating e(X)e(X). The motivation is propensity sufficiency: under unconfoundedness, the propensity score is a balancing score, so any representation Φ that preserves e(X)e(X) retains all information needed for CATE estimation. Outcome-only training tends to discard propensity signal in pursuit of marginal-likelihood fit, biasing τ^=μ^1−μ^0 τ= μ_1- μ_0 toward the dominant treatment arm. Jointly optimizing the propensity head therefore acts as an architectural regularizer on Φ , and in practice yields markedly more stable CATE estimates (Shi et al., 2019). We extend DragonNet with three key enhancements for recommender systems. First, we feed sequential user-event features through a transformer layer (Section 3) before they enter Φ , letting the encoder absorb temporal context such as recency and engagement cadence. Second, we extend the outcome heads to a multi-outcome configuration using learned product embeddings, so the model can score newly introduced products without adding a product-specific output head. Third, and more importantly, we couple this causal predictor with the neural bandit in Section 2.2, forming a DragonBandit policy that explores over incremental action values rather than predicted responses. 2.2. Neural Bandit Exploration Reliance on logged data without any mechanism for exploration reinforces feedback loops, leading to “rich-get-richer” dynamics and suboptimal long-term performance (Li et al., 2010; Su et al., 2024; Swaminathan and Joachims, 2015; Nguyen et al., 2026). For causal modeling, exploration is additionally important to improve overlap and reduce dependence on the incumbent logging policy (Schnabel et al., 2016; Kasy and Sautmann, 2021), supporting causal identification where the required assumptions hold (Neal, 2020). At each round, DragonBandit draws posterior incremental scores rather than passing deterministic point estimates to the allocator in Section 2.3. The multi-turn simulation examines this mechanism under feedback-loop bias. Let hth_t denote the history and remaining constraint capacity at round t, t(ht)C_t(h_t) the corresponding feasible set, and xt∗(τ~t)x_t^*( τ_t) its sampled-score solution. For nonzero exploration, the sampled logits have a nondegenerate Gaussian distribution; because the sigmoid heads are monotone, every open score ordering has positive probability. Therefore, (2) πα(i∣u,ht)=Pr[xu,i,t∗(τ~t)=1∣ht]>0 _α(i u,h_t)= \! [x^*_u,i,t( τ_t)=1 h_t ]>0 for every (u,i)(u,i) contained in some xt∈t(ht)x_t _t(h_t). Thus exploration guarantees positivity conditional on the feasible action. 2.2.1. Neural Thompson sampling via Laplace approximation We approximate the posterior over network parameters by a Gaussian centered at the MAP solution θ^MAP θ_MAP via the linearized Laplace approximation (LLA) (Su et al., 2024; Daxberger et al., 2021; Foong et al., 2019; Raha et al., ). The LLA computes the Gauss–Newton curvature of the log-likelihood at θ^MAP θ_MAP and propagates it through the network’s local linearization, yielding a closed-form Gaussian on the logit fθ(x∗)f_θ(x^*) with mean fθ^(x∗)f_ θ(x^*) and variance σ02+g(x∗)⊤Ω−1g(x∗) _0^2+g(x^*) ^-1g(x^*), where g(x∗)g(x^*) is the Jacobian and Ω is the curvature matrix. Thompson sampling proceeds by drawing a logit sample and applying σ(⋅)σ(·), so the same supervised checkpoint can be deployed with or without exploration and no retraining is required. Production scale is achieved through lightweight curvature approximations (low-rank, diagonal, or last-layer variants (Nilsen et al., 2022; Zhang et al., 2020; Riquelme et al., 2018)); we use the last-layer variant for our pretrained DragonNet. We also evaluated Bayes by Backprop (Appendix B.1) offline, but use LLA in production because it adds uncertainty sampling to the existing checkpoint with substantially lower training and serving overhead. 2.3. Large-scale Allocation with Constraints The neural bandit improves decisions across rounds, but each round must still coordinate assignments under a common set of business constraints. Let xu,i,tx_u,i,t indicate whether item i∈ℐi is allocated to user u∈u at round t=1,…,Tt=1,…,T. Using the sampled incremental objective τ~u,i,tobj τ_u,i,t obj from DragonBandit, the horizon objective is (3) maxxu,i,t max_x_u,i,t ∑t=1T∑u,iτ~u,i,tobjxu,i,t, _t=1^T _u,i τ_u,i,t objx_u,i,t, s.t. ∑u,iτu,i,tguardrailkxu,i,t≤Cguardrailk,∀t,k, _u,i _u,i,t guardrail_kx_u,i,t≤ C_guardrail_k, ∀\,t,k, ∑ixu,i,t≤Cfcap,∀u,t, _ix_u,i,t≤ C_fcap, ∀\,u,t, xu,i,t∈0,1,∀u,i,t. x_u,i,t∈\0,1\, ∀\,u,i,t. Here τu,i,ty=fty(u,i,1)−fty(u,i,0) _u,i,t^y=f_t^y(u,i,1)-f_t^y(u,i,0) is the round-specific sampled treatment effect on metric y. We use a sequential policy that solves one constrained optimization problem per round using the current posterior sample from DragonBandit. After each round, DragonBandit is updated using the observed feedback before producing the posterior sample for the next round. In the absence of shared allocation constraints, the resulting policy is equivalent to Thompson sampling: for each user, it selects the feasible action that maximizes the sampled incremental reward. 2.3.1. Scalability via Dual Decomposition At production scale, each round contains tens of millions of users and hundreds of items, so its batch has ||×|ℐ||U|×|I| variables and is intractable for general-purpose solvers. Suppressing the round index, we relax xu,i,tx_u,i,t to represent an action probability and solve the resulting large-scale problem using a smoothed dual-decomposition method (Basu et al., 2020). The method adds a small ridge perturbation γ2‖x‖2 γ2\|x\|^2 to the primal and dualizes the K global constraints with multipliers λ≥0λ≥ 0; by Danskin’s theorem, the resulting dual gγ(λ)g_γ(λ) is differentiable with Lipschitz gradient ∇gγ(λ)=Axγ∗(λ)−b∇ g_γ(λ)=A\,x_γ^*(λ)-b, where the primal minimizer decomposes per user: (4) xγ,u∗(λ)=Πu[−1γ(AuTλ+cu)],x_γ,u^*(λ)= _C_u\! [- 1γ (A_u^Tλ+c_u ) ], with Πu _C_u an O(|ℐ|log|ℐ|)O(|I| |I|) projection onto the per-user frequency-cap polytope. The solver then maximizes gγg_γ over the K-dimensional dual with Nesterov-accelerated ascent, giving per-iteration cost linear in ||⋅|ℐ||U|·|I| versus O((|||ℐ|)3.5)O((|U||I|)^3.5) for interior-point methods. The regularization γ is picked as the largest value satisfying γx^Tx^2|cTx^|<10−3 γ\, x^T x2\,|c^T x|<10^-3, so the ridge perturbation contributes <0.1%<0.1\% of the objective and the perturbed optimum is practically indistinguishable from the true LP optimum; when A is ill-conditioned across constraint scales, we apply Jacobi row preconditioning. In steady state, we warm-start the dual from the previous period’s λ∗λ^*, which under stable input distributions (KS-tested) achieves over 99% of the current optimum and also serves as an SLA fallback when the solver does not converge in time. 3. Incrementality Transformer Architecture 3.1. Temporal Features We construct a touchpoint sequence S=(s1,…,sL)S=(s_1,…,s_L) over a lookback window, where each sis_i encodes a (channel, action) interaction such as (AD, IMPRESSION) or (EMAIL, OPEN); we also inject prior conversion events into the sequence. Each token is mapped through a learned embedding matrix ET∈ℝV×dE_T ^V× d, with V interaction types plus a leading [CLS] token (BERT-style (Devlin et al., 2019)) that summarizes the sequence even when empty. We use the temporal encoding described below. Temporal position. Member interactions occur on irregular schedules, so each token combines its interaction embedding with a day-since embedding EDAYSE_DAYS, day-of-week embedding EDOWE_DOW, and the small-embedding-friendly sinusoidal positional encoding PEiPE_i of Foumani et al. (2023): (5) Ei=ET(si)+EDAYS(ti)+EDOW(dowi)+PEi.E_i=E_T(s_i)+E_DAYS(t_i)+E_DOW(dow_i)+PE_i. This represents both sequence order and elapsed time without expanding the interaction vocabulary. Appendix B.2 gives the sinusoidal parameterization. Multi-head attention. We apply standard multi-head self-attention (Vaswani et al., 2017) with Q=K=V=(E1,…,EL)Q=K=V=(E_1,…,E_L), and pool by taking only the [CLS] output as the shared representation ES∈ℝdE_S ^d, which is concatenated with member embeddings and dense features before Φ . 3.2. Multi-Outcome Extension Marketing campaigns generate multiple outcomes (products) per member, so we extend the single-outcome DragonNet to a multi-label setting YkY_k, k∈1,…,Kk∈\1,…,K\. Because per-product conversion rates are highly imbalanced, we use inverse-frequency weights ckc_k and combine the standard BCE loss (with sigmoid σ) into a per-head outcome loss (6) ℒY|T=t=1K∑kck⋅BCE(yk,y^k),ℒY=ℒY|T=1+ℒY|T=0,L_Y|T=t= 1K\! _kc_k·BCE(y_k, y_k), _Y=L_Y|T=1+L_Y|T=0, counting each head’s loss only when the corresponding T is observed (as in DragonNet/TARNET). Treatment prediction uses standard BCE on the propensity-head output e^(X) e(X), ℒT=[BCE(T,e^(X))]L_T=E[BCE(T, e(X))]. Following Shi et al. (2019), we add targeted regularization based on the efficient influence function (EIF) for the ATE under unconfoundedness (Chernozhukov et al., 2018): (7) ϕEIF=τ^(X)+Te^(X)(Y−μ^1)−1−T1−e^(X)(Y−μ^0)−ATE. _EIF= τ(X)+ T e(X)(Y- μ_1)- 1-T1- e(X)(Y- μ_0)-ATE. The EIF is Neyman-orthogonal. We use its one-step correction as a training regularizer, not as a substitute for identification assumptions or as a stand-alone guarantee of unbiased CATE estimation: (8) ℒtarreg=[‖Y−(Y^+ϵψ)‖2],L_tarreg=E\! [ \|Y-( Y+ε\,ψ) \|^2 ], with ψ=T/e^(X)−(1−T)/(1−e^(X))ψ=T/ e(X)-(1-T)/(1- e(X)) (the TMLE clever covariate (van der Laan and Rubin, 2006)), learnable scalar ϵε, and Y Y the prediction from the head matching observed T. 3.3. Outcome Embeddings New-product launches create a cold-start problem: marketers want to promote an offering before the model has seen examples for it. Our use of “CLIP-inspired” refers narrowly to a normalized shared embedding space with temperature-scaled similarity; it does not use CLIP’s large-scale contrastive pretraining. Instead of aligning images and text, we align a member’s outcome state with an attribute-derived product representation, allowing the same member encoder to produce scores for products whose attributes can be embedded, including products unseen during training. Concretely, each outcome tower (a two-dense-block MLP with linear projection) drops the final sigmoid and returns hidden logits ht(X),hnt(X)∈ℝdoh_t(X),h_nt(X) ^d_o representing the member’s expected state under T=1T=1 and T=0T=0. A single shared outcome embedding matrix EO∈ℝK×doE_O ^K× d_o encodes the K products and is consumed by both towers, reflecting the fact that the product’s intrinsic semantics are invariant to treatment. 3.3.1. Outcome Autoencoder EOE_O is the bottleneck of a lightweight autoencoder EO=Encoder(EOinput)E_O=Encoder(E^input_O) over outcome attribute representations. EOinputE^input_O is constructed by mapping each product attribute through a learned embedding and concatenating; attributes can be either structured (one-hot product taxonomy, business-line, format) or unstructured (LLM-derived embeddings of product descriptions). The autoencoder is trained jointly with the main task using MSE reconstruction (9) ℒrecon=1K∑k‖EO,kinput−Decoder(EO,k)‖2.L_recon= 1K\! _k \|E^input_O,k-Decoder(E_O,k) \|^2. This serves as semantic regularization: compared with a directly trained, task-specific EOE_O, the reconstruction objective encourages EOE_O to retain the geometry of the input attribute space. The same setup enables large-scale embedding-space simulation, since pre-computed ht(X)h_t(X) and hnt(X)h_nt(X) can be paired with arbitrary EOE_O samples without re-running the full model. 3.3.2. Outcome Matrix Layer The final outcome logits are L2-normalized dot products with learnable log-inverse-temperatures νt,νnt _t, _nt (initialized to log14.0 14.0 following CLIP): (10) y^t=ht(X)‖ht(X)‖⋅EOT‖EO‖⋅exp(νt),y^nt=hnt(X)‖hnt(X)‖⋅EOT‖EO‖⋅exp(νnt). y_t= h_t(X)\|h_t(X)\|· E_O^T\|E_O\|· ( _t), y_nt= h_nt(X)\|h_nt(X)\|· E_O^T\|E_O\|· ( _nt). Each row of EOE_O is L2-normalized so that similarity depends purely on the angle between member-state and product representations, keeping logits on a consistent scale across products and treatment arms. 3.4. Complete Loss Function The complete training objective sums the outcome, treatment, targeted-regularization, and reconstruction losses, (11) ℒ=ℒY|T=1+ℒY|T=0+ℒT+ℒtarreg+λreconℒrecon,L=L_Y|T=1+L_Y|T=0+L_T+L_tarreg+ _reconL_recon, where λrecon _recon weights the reconstruction loss; ℒreconL_recon is included only when outcome attribute encodings are provided. Although the complete research architecture contains several losses and optional modules, the serving path is modular rather than a jointly tuned monolith: outcome embeddings are used only for multi-product scoring, LLA is applied after supervised training to the last layer, and the LP consumes exported scores independently of model training. This separation lets each module be disabled or validated without retraining the rest of the decision pipeline, while the remaining tuning burden and feature sensitivity are limitations of the current shared-representation model. 4. Offline Simulations In order to showcase the strengths of our proposed methodology, we perform an offline simulation study with a publicly available dataset. We focus on comparing the following approaches to targeting recommendations. • Incremental Modeling + Constrained Optimization (ours): integrates uplift modeling with constrained optimization for system-level incremental targeting; uses dual decomposition (Basu et al., 2020) at production scale and OR-Tools for simulation studies. • Bandit Incremental Modeling + Constrained Optimization: extends our method with BNN-based exploration to address feedback-loop bias (Nguyen et al., 2026); same solver setup as above. • Propensity Modeling + Constrained Optimization: optimizes predicted scores under constraints (Agarwal et al., 2015; Makhijani et al., 2019) without modeling organic outcomes. • Incremental Modeling + Ranking: ranks by estimated uplift (Chen et al., 2020; Meng et al., 2025; Sun et al., 2024) but ignores global constraints. • Propensity Modeling + Ranking: ranks by predicted response probability (Li et al., 2010; Covington et al., 2016; Cheng et al., 2016; Ying et al., 2018). 4.1. Dataset We use the Open Bandit Dataset (OBD) (Saito et al., 2020), a real-world logged bandit dataset. We utilize the random-policy subset, synthetically mapping the original 34 products that were recommended to >400>400K users to 5 distinct actions that are relevant to the production incrementality use case: recommendation to one of four business lines or no-recommendation. The no-recommendation action is a key difference between incremental and non-incremental targeting. We consider the logged reward within the OBD dataset as indicative of a conversion event to one of the four business lines. We augment that reward by coupling it with the average product price of the four business lines and apply per-business-line LP volume bounds, as shown in Table 1. Table 1. Average product price and LP constraint bounds (as % of audience) for each business line. Business Line Avg. Price Min. Vol. Max. Vol. A $5 5% 10% B $10 5% 30% C $10 5% 30% D $200 30% 50% We also constructed a treatment variable that indicates whether a member was exposed to a marketing campaign showcasing one of the four business lines. This setup mirrors the production ecosystem and allows us to demonstrate the importance of incremental modeling. Finally, we assign a cost of $0.1 to each recommendation/targeting action to capture operational and bidding costs. 4.2. Setup We performed an 80/20 split of the full dataset into training and prediction sets. We trained the predictive models (incremental and propensity) on the training set, generating incremental/propensity scores for the prediction set. We applied constrained optimization or ranking to arrive at the final recommendations for the prediction set (one of the five actions described above), with the objective function constructed by coupling the incremental/propensity scores for conversion to each business line with the corresponding average product price. We solve the offline simulation LP with Google OR-Tools, which is tractable at this scale (∼400 400K members, 5 actions) and convenient for reproduction; at full production scale we use the dual-decomposition method described in Section 2.3 instead. We apply the volume bounds from Table 1. The ranking baseline assigns each member to the action with the highest predictive score, disregarding global business constraints. We consider two simulation regimes: a single-turn (static) evaluation and a multi-turn (online) evaluation. For the single-turn setting, we construct the training log from a random logging policy in which all actions are approximately equally represented, and we evaluate each method once on the full prediction set using fixed model scores. For the multi-turn setting, we intentionally bias the training log by severely under-sampling a subset of actions, so that their estimated uplifts become high-variance and can even exhibit sign errors relative to the OBD ground-truth uplifts. We then simulate deployment over T=200T=200 rounds using the prediction set as the environment: at each round, each method selects actions for the current batch, observes the realized rewards from its own recommendations, updates its training data accordingly, and incrementally updates the model before the next round. 4.3. Results 4.3.1. Single-turn evaluation We evaluated the different methods by calculating the corresponding average rewards with Doubly Robust Policy Evaluation. As shown in Table 2, incremental scores lead to higher rewards compared to propensity scores. We also provide the corresponding send volumes in Table 3, where it can be seen that incremental targeting leads to a higher percentage of no-recommendations as the engine is able to identify members likely to convert organically. Table 2. Average Rewards, Costs, and Net Returns ML Model Optim. Avg Reward Avg Cost Avg Net Return Incremental Constr. Opt. $0.55±0.14 0.55± 0.14 $0.091±0.001 0.091± 0.001 $0.46±0.13 0.46± 0.13 Propensity Constr. Opt. $0.49±0.18 0.49± 0.18 $0.1±0 0.1± 0 $0.39±0.18 0.39± 0.18 Incremental Ranking $0.46±0.15 0.46± 0.15 $0.090±0.001 0.090± 0.001 $0.37±0.15 0.37± 0.15 Propensity Ranking $0.41±0.20 0.41± 0.20 $0.1±0 0.1± 0 $0.31±0.20 0.31± 0.20 Table 3. Send Volumes Method Bus. Line A Bus. Line B Bus. Line C Bus. Line D No Rec. Incremental + Constr. Opt. 5%5\% 30%30\% 26%26\% 30%30\% 9%9\% Propensity + Constr. Opt. 5%5\% 18%18\% 27%27\% 50%50\% 0%0\% Incremental + Ranking 0%0\% 46%46\% 44%44\% 0%0\% 10%10\% Propensity + Ranking 0%0\% 0%0\% 0%0\% 100%100\% 0%0\% The send volumes also highlight the difference between Constrained Optimization and Ranking. As expected, Ranking provides solutions that deviate significantly from the desired business constraints. We show the average costs in Table 2, where it can be seen that incremental modeling leads to reduced cost due to the reduced send volumes described above. The corresponding net returns are shown in Table 2, where Incremental Modeling coupled with Constrained Optimization provides the highest overall performance. 4.3.2. Multi-turn evaluation The previous single-turn evaluation shows that only approaches with Constrained Optimization can adequately satisfy the volume constraint bounds induced by business constraints, so we restrict attention to these approaches in the multi-turn evaluation. This multi-turn evaluation demonstrates the advantage of having an explicit mechanism to balance exploration and exploitation at the system level. The initial training-data bias we inject is designed to mirror common challenges in industry applications, such as cold-start, low representation for certain actions, and non-stationarity where the true reward function evolves over time (e.g., an initially low-performing action later becomes high-performing). Figure 2(a) shows that the Bandit Incremental Model is able to overcome this bias and ultimately outperform the other approaches after learning from the outcomes of its own actions, even though it may underperform its greedy counterpart in the first few steps of online learning. This reflects the short-term cost of exploration in exchange for long-term gains: after roughly 50 model updates, the Bandit Incremental Model begins to outperform both greedy variants. Finally, we show that the benefit of exploration is more pronounced when the initial training-data bias is larger, as illustrated in Figure 2(b). (a) Small bias setting. (b) Large bias setting. Figure 2. Average cumulative reward and 95% confidence intervals after 200 rounds of feedback. The confidence intervals are calculated from 30 simulation runs. 4.4. Ablation Study We tested 8 configurations on a fixed train/validation snapshot, each repeated 5 times with common hyperparameters. Figure 3(a) summarizes outcome, treatment, and uplift performance. Adding outcome embeddings preserves outcome AUROC and produces the highest uplift AUUC point estimate. Removing dense features reduces outcome AUROC, while its uplift effect depends on the configuration: AUUC increases for the base model but decreases slightly when bandit exploration is enabled. This is consistent with the dense features being prognostic rather than treatment-effect modifiers. Appendix A provides the full methodology, results, and outcome-embedding diagnostic. (a) Outcome prediction AUROC. (b) Treatment prediction AUROC. (c) Uplift AUUC. Figure 3. Ablation results. Error bars denote 95% confidence intervals from 5 training runs. 5. Production Deployment and Experimentation We now turn from modeling to production. Deploying causal targeting at scale taught us that a handful of pragmatic enhancements beyond the core model and optimizer are what make the system work in practice. The most important are the construction of training data and the control of spend and delivery. These enhancements are largely invisible to offline simulations and ablation studies, which hold the data-generating process and delivery pipeline fixed, yet in our experience they are important for real-world success. We describe these enhancements first, followed by the serving architecture, agentic experimentation setup, and online results. 5.1. Training-Data Construction for Causal Estimation Production logs allow us to build sequences of marketing touches and conversions, but causal estimation requires careful construction of the context sequence X, treatment action T, and conversion label Y. The context must contain only information available before treatment, while treatment and outcome are assigned from subsequent non-overlapping windows. Let R denote the snapshot run date. For each member, production samples D=R−(WC+WT+WD)+UD=R-(W_C+W_T+W_D)+U, where U∼Uniform0,…,WD−1U \0,…,W_D-1\, WD=90W_D=90 days, WT=7W_T=7 days, and WC=30W_C=30 days. Thus D∈[R−127,R−38]D∈[R-127,R-38]. The model input contains marketing interactions and prior conversions in [D−60,D)[D-60,D); post-D events are excluded. We set T=1T=1 when at least one qualifying email send, on-platform impression, or video view occurs in [D,D+7)[D,D+7). We set Y=1Y=1 when the corresponding business-line or product-family conversion occurs in [D+7,R][D+7,R], and Y=0Y=0 otherwise. This reserves at least 30 days of follow-up while earlier dates have longer outcome windows. Randomizing D avoids a last-touch label, captures long-term action effects, and preserves variable-length histories. 5.2. Cost and Delivery Control Controlling how much the causal policy spends and delivers matters for two reasons: a standing business requirement to pace committed budgets over the fiscal quarter, and an experimentation requirement to compare arms at matched delivery. Both stem from the same underlying gap between an assignment and its realized delivery. At a given round, with t suppressed, the LP in (3) chooses assignments xu,ix_u,i, but an assignment is not a guaranteed delivery in an auction-mediated, activity-gated channel. A member may not return during the campaign window, or the send may lose the downstream auction. We therefore estimate the probability of delivery, (12) pu,i=P(delivered within window∣u,i),p_u,i=P(delivered within window u,i), using a classifier with isotonic calibration per member-lifecycle segment. The estimate enters the LP through the expected-impression constraint ∑u,ipu,ixu,i∈[Cimpl,Cimpu] _u,ip_u,ix_u,i∈[C^l_imp,C^u_imp] and the expected-cost constraint ∑u,ic^u,ixu,i∈[Ccostl,Ccostu] _u,i c_u,ix_u,i∈[C^l_cost,C^u_cost]. Their bounds are calibrated to the observed BAU delivery envelope. This prevents the optimizer from concentrating assignments on members who appear incremental but are difficult or expensive to reach. Across multiple rounds, such as in quarterly budget pacing or cumulative spend matching in an experiment, realized delivery and cost may differ from their LP predictions. We therefore introduce a feedback controller that adjusts the round-level cost target. Let B be the committed budget, StS_t the cumulative realized spend, and q(t/H)q(t/H) the desired cumulative pacing curve over a horizon H, with q(0)=0q(0)=0 and q(1)=1q(1)=1. The outer loop computes (13) St⋆=Bq(t/H),et=St⋆−St.S_t =Bq(t/H), e_t=S_t -S_t. Let DtD_t denote the realized spend during interval t, and let At=∑u,ic^u,i,txu,i,tA_t= _u,i c_u,i,tx_u,i,t denote the LP-predicted spend of the selected assignments. The inner loop estimates the spend-realization ratio as rt=Dt/Atr_t=D_t/A_t and smooths it using r^t=αrt+(1−α)r^t−1 r_t=α r_t+(1-α) r_t-1, where α∈(0,1]α∈(0,1] weights the newest observation. It then updates the LP cost target: (14) Ccost,t+1 C_cost,t+1 =Π[Cmin,Cmax](Ccost,t+κetmax(r^t,ϵ)), = _[C_ ,C_ ] (C_cost,t+κ e_t ( r_t,ε) ), where κ>0κ>0 is the controller gain. The target Ccost,tC_cost,t can be converted into a two-sided LP constraint using a predefined tolerance around the target. When realized spend falls behind the pacing curve, the controller raises the assignment cap in proportion to the gap and estimated deliverability; when spend runs ahead, it lowers the cap. The projection preserves operational bounds. An early A/B test run without the above controls showed the treatment arm under-delivering: the causal policy withholds and reallocates sends, lowering treatment impressions relative to control. A raw total-bookings comparison then penalizes the treatment for delivering less rather than for choosing worse, confounding policy quality with delivery volume. Constraining treatment to the BAU impression and cost envelope equalizes delivery across arms and restores a clean, like-for-like read. 5.3. Serving Architecture The system is served as a batch pipeline. DragonBandit scores μ1 _1, μ0 _0, and e for the eligible population and samples incremental objectives via last-layer linearized Laplace Thompson sampling (Section 2.2); a dual-decomposition solver (Basu et al., 2020) computes the constrained allocation (3) under global guardrails, warm-started from the previous period’s dual λ⋆λ ; and the feedback controller (Section 5.2) adjusts budget caps from realized deliverability before assignments are materialized to the campaign-management system. Because scoring and allocation are decoupled from serving, the same supervised checkpoint is deployed with or without exploration, and non-convergence solver falls back to the warm-started dual, which recovers over 99% of the current optimum under stable input distributions within the SLA. 5.4. Agentic Experimentation We evaluated the end-to-end framework with an eight-week online A/B test on LinkedIn Feed marketing traffic, against the deployed business-as-usual (BAU) targeting stack. BAU is a standard two-tier recommender: the retrieval tier pre-selects each campaign’s audience using a propensity-based scoring model together with marketer-defined criteria, and the ranking tier is a second propensity-based model that estimates engagement probability and is followed by ranking heuristics for final assignment. The treatment arm keeps the marketer-defined criteria in retrieval and replaces both propensity layers with a single decision layer that scores members by predicted causal uplift across all campaigns and assigns them via the LP under global constraints. A meaningful policy difference is that the treatment arm can withhold a send whenever the predicted incremental value is negative or no feasible positive-incremental option exists. Members were randomly assigned 50/50 to the two experiment arms. Standing up a representative BAU control and a matched treatment arm at scale requires configuring hundreds of live campaigns across products, segments, and budgets. Manual configuration is infeasible and, we found, subject to two subtle failure modes that bias measurement yet are invisible offline. First, when out-of-scope campaigns are suppressed via an exclusion segment defined by dynamic criteria (company, locale, activity), members drift across arms in a way correlated with platform activity, contaminating the intent-to-treat contrast. Second, when arms are assembled from shared segment definitions, a control audience can silently inherit a treatment send decision through a reused sub-segment. We therefore built an agentic campaign-setup tool that duplicates campaigns into hundreds of budget-split variants, freezes each audience to a static snapshot for the experiment’s duration, audits segment lineage so that no control segment depends on a treatment-derived decision, and materializes campaigns as drafts promoted by a single activation toggle with symmetric rollback. This provides identical scaffolding across arms, so the only systematic difference is the decision layer. It also improves operational safety. Budget parity is enforced by setting caps proportional to audience sizes and recalibrating as arm sizes drift. 5.5. Results Measured against long-term-value metrics over an eight-week period, the treatment arm achieved a statistically significant +7.20%+7.20\% lift (p=0.041p=0.041, 95% CI: [0.31%,14.09%][0.31\%,14.09\%]). The experiment evaluates the end-to-end policy as deployed. Its components satisfy coupled system requirements rather than representing independent product features: causal scoring defines the incremental objective, the exploration policy provides stochastic treatment variation that improves positivity (overlap), the LP enforces global business constraints, and the withhold rule prevents assignments with negative incremental value. Removing any of these components changes the operating requirements or the data-generating policy. The online result therefore measures the system-level impact of the complete production policy, while the offline studies examine individual mechanisms under controlled settings. Appendix A Detailed Ablation Study A.1. Protocol We ran targeted ablation simulations on a static train/validation snapshot, holding hyperparameters (learning rate, epochs, batch size, hidden sizes) constant across configurations and varying only the indicated change. We tested 8 configurations, each repeated 5 times. The base model is the architecture from Section 3 on top of the DragonNet causal head, without bandit sampling or outcome embeddings (OE); we then added bandit and OE separately, and additionally removed three input categories: dense features, entity embeddings, and sequential interactions. These are not exhaustive component ablations: in particular, they do not separately isolate the Transformer operator from its sequence input, targeted regularization, or every interaction among loss terms. A.2. Outcome and Treatment Metrics Outcome ROC-AUC is computed against the common validation dataset using the observed treatment for masking, y^=Ty^t+(1−T)y^nt y=T y_t+(1-T) y_nt, and macro-averaged across labels. Treatment AUROC evaluates the propensity head against the observed treatment assignment. A.3. Uplift Results Following CausalML (Chen et al., 2020), we rank observations by predicted uplift and compute (15) CumLift(k)=∑i=1kYiTi∑i=1kTi+ϵ−∑i=1kYi(1−Ti)∑i=1k(1−Ti)+ϵ,CumLift(k)= _i=1^kY_iT_i _i=1^kT_i+ε- _i=1^kY_i(1-T_i) _i=1^k(1-T_i)+ε, with CumGain(k)=kCumLift(k)CumGain(k)=k\,CumLift(k) and normalized area under the gain curve as AUUC. Implementations of AUUC and Qini differ (Gutierrez and Gérardy, 2017; Radcliffe, 2007; Devriendt et al., 2020); we use the CausalML implementation. Removing dense features is neutral or beneficial for uplift AUUC despite reducing outcome AUROC from 0.857 to 0.826 and minimally affecting treatment AUROC. A plausible explanation is that these features are prognostic and predict outcomes regardless of treatment, rather than serving as treatment-effect modifiers. They may therefore add variance when the shared DragonNet representation must serve both propensity and outcome losses. This remains a hypothesis rather than an automated feature-selection mechanism; architectures such as FlexTENet (Curth and van der Schaar, 2021) could explicitly separate the two subspaces. A.4. Outcome-Embedding Diagnostic We train on 9 of 10 outcomes, draw 5,000 hypothetical products from the learned embedding space, and score each against the held-out 10th product across the validation members. Figure 4 shows that samples near semantically similar trained products have higher AUROC against the held-out target, indicating that the geometry retains outcome-relevant structure. Because this diagnostic uses one held-out product and sampled embeddings rather than prospective real launches, it does not establish zero-shot production effectiveness. Figure 4. Outcome-embedding PCA. Each sampled hypothetical product is colored by AUROC z-score against the held-out product. Appendix B Additional Architectural Details B.1. Bayes by Backprop Bayes by Backprop (Blundell et al., 2015; Kingma et al., 2015) learns a Gaussian variational posterior qϕ(θ)q_φ(θ) by maximizing the evidence lower bound. Predictive probabilities are approximated by Monte Carlo sampling, (16) P(y∗=1∣x∗)≈1M∑m=1Mσ(fθ(m)(x∗)),θ(m)∼qϕ.P(y^*=1 x^*)≈ 1M _m=1^Mσ(f_θ^(m)(x^*)), θ^(m) q_φ. Compared with LLA, it captures non-local posterior uncertainty but adds stochastic training and repeated inference-time sampling. For a minibatch variant, an additional βℒKL _KL term regularizes the posterior toward the prior; β=1β=1 gives the standard ELBO objective. We did not explore alternative KL-weight schedules in the reported experiments. B.2. Temporal Position Encoding For sequence position pospos and embedding coordinate i, we use the small-embedding-friendly sinusoidal variant (17) PE(pos,2i)=sin(posωinew),PE(pos,2i+1)=cos(posωinew),PE_(pos,2i)= (pos\, _i^new), PE_(pos,2i+1)= (pos\, _i^new), where ωinew=ωid/L _i^new= _id/L and ωi=10000−2i/d _i=10000^-2i/d. The scaling adapts the standard sinusoid to the short interaction vocabulary and sequence length used in our setting. Ethical Considerations We identify no ethical implications specific to the proposed optimization method beyond the standard privacy, fairness, and experimentation considerations of large-scale recommendation systems; deployments should follow applicable data-governance and review processes. References D. Agarwal, S. Chatterjee, Y. Yang, and L. Zhang (2015) Constrained optimization for homepage relevance. In Proceedings of the 24th International Conference on World Wide Web, p. 375–384. Cited by: §1, 3rd item. K. Basu, A. Ghoting, R. Mazumder, and Y. Pan (2020) ECLIPSE: an extreme-scale linear program solver for web-applications. External Links: 2007.15936, Link Cited by: §2.3.1, 1st item, §5.3. D. Bertsimas, D. B. Brown, and C. Caramanis (2011) The theory of robust optimization. SIAM Review. Cited by: §1. C. Blundell, J. Cornebise, K. Kavukcuoglu, and D. Wierstra (2015) Weight uncertainty in neural networks. External Links: 1505.05424, Link Cited by: §B.1. S. Bonner and F. Vasile (2018) Causal embeddings for recommendation. In Proceedings of the 12th ACM Conference on Recommender Systems, Cited by: §1. H. Chen, T. Harinen, J. Lee, M. Yung, and Z. Zhao (2020) CausalML: python package for causal machine learning. External Links: 2002.11631 Cited by: §A.3, 4th item. J. Chen, W. Wenjie, C. Gao, P. Wu, J. Wei, and Q. Hua (2024) Treatment effect estimation for user interest exploration on recommender systems. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, p. 1861–1871. Cited by: §1.1. H. Cheng, L. Koc, J. Harmsen, T. Shaked, T. Chandra, H. Aradhye, G. Anderson, G. Corrado, W. Chai, M. Ispir, R. Anil, Z. Haque, L. Hong, V. Jain, X. Liu, and H. Shah (2016) Wide & deep learning for recommender systems. In Proceedings of the 1st Workshop on Deep Learning for Recommender Systems (DLRS 2016), p. 7–10. Cited by: 5th item. V. Chernozhukov, D. Chetverikov, M. Demirer, E. Duflo, C. Hansen, W. Newey, and J. Robins (2018) Double/debiased machine learning for treatment and structural parameters. The Econometrics Journal 21 (1), p. C1–C68. External Links: ISSN 1368-4221, Document, Link Cited by: §2.1, §3.2. P. Covington, J. Adams, and E. Sargin (2016) Deep neural networks for YouTube recommendations. In Proceedings of the 10th ACM Conference on Recommender Systems (RecSys ’16), p. 191–198. Cited by: 5th item. A. Curth and M. van der Schaar (2021) On inductive biases for heterogeneous treatment effect estimation. External Links: 2106.03765, Link Cited by: §A.3. E. Daxberger, A. Kristiadi, A. Immer, R. Eschenhagen, M. Bauer, and P. Hennig (2021) Laplace redux-effortless bayesian deep learning. Advances in neural information processing systems 34, p. 20089–20103. Cited by: §2.2.1. J. Devlin, M. Chang, K. Lee, and K. Toutanova (2019) BERT: pre-training of deep bidirectional transformers for language understanding. External Links: 1810.04805, Link Cited by: §3.1. F. Devriendt, T. Guns, and W. Verbeke (2020) Learning to rank for uplift modeling. CoRR abs/2002.05897. External Links: Link, 2002.05897 Cited by: §A.3. A. Y. Foong, Y. Li, J. M. Hernández-Lobato, and R. E. Turner (2019) ’In-between’uncertainty in bayesian neural networks. arXiv preprint arXiv:1906.11537. Cited by: §2.2.1. N. M. Foumani, C. W. Tan, G. I. Webb, and M. Salehi (2023) Improving position encoding of transformers for multivariate time series classification. Data Mining and Knowledge Discovery 38 (1), p. 22–48. External Links: ISSN 1573-756X, Link, Document Cited by: §3.1. C. Gao, Y. Zheng, W. Wang, F. Feng, X. He, and Y. Li (2024) Causal inference in recommender systems: a survey and future directions. ACM Transactions on Information Systems 42 (4), p. 1–32. Cited by: §1. P. Gutierrez and J. Gérardy (2017) Causal inference and uplift modelling: a review of the literature. In Proceedings of The 3rd International Conference on Predictive Applications and APIs, C. Hardgrove, L. Dorard, K. Thompson, and F. Douetteau (Eds.), Proceedings of Machine Learning Research, Vol. 67, p. 1–13. External Links: Link Cited by: §A.3. M. Kasy and A. Sautmann (2021) Adaptive treatment assignment in experiments for policy choice. Econometrica 89 (1), p. 113–132. Cited by: §2.2. D. P. Kingma, T. Salimans, and M. Welling (2015) Variational dropout and the local reparameterization trick. Advances in neural information processing systems 28. Cited by: §B.1. S. R. Künzel, J. S. Sekhon, P. J. Bickel, and B. Yu (2019) Metalearners for estimating heterogeneous treatment effects using machine learning. Proceedings of the National Academy of Sciences 116 (10), p. 4156–4165. External Links: ISSN 1091-6490, Link, Document Cited by: §2.1. L. Li, W. Chu, J. Langford, and R. E. Schapire (2010) A contextual-bandit approach to personalized news article recommendation. In Proceedings of the 19th international conference on World wide web, p. 661–670. Cited by: §2.2, 5th item. R. Makhijani, S. Chakrabarti, D. Struble, and Y. Liu (2019) LORE: a large-scale offer recommendation engine with eligibility and capacity constraints. In Proceedings of the 13th ACM Conference on Recommender Systems, p. 160–168. Cited by: §1, 3rd item. J. McInerney, B. Brost, P. Chandar, R. Mehrotra, and B. Carterette (2020) Counterfactual evaluation of slate recommendations with sequential reward interactions. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, p. 1779–1788. Cited by: §1. C. Meng, C. Zhai, X. Wang, S. Liu, X. Feng, L. Hu, X. Li, H. Li, and K. Gai (2025) Enhancing online video recommendation via a coarse-to-fine dynamic uplift modeling framework. In Proceedings of the 19th ACM Conference on Recommender Systems, p. 82–92. Cited by: §1.1, 4th item. B. Neal (2020) Introduction to causal inference. Course lecture notes (draft) 132. Cited by: §2.2. P. Nguyen, B. Zelditch, J. Chen, R. Patra, and C. Wei (2026) BanditLP: large-scale stochastic optimization for personalized recommendations. arXiv preprint arXiv:2601.15552. Cited by: §2.2, 2nd item. G. K. Nilsen, A. Z. Munthe-Kaas, H. J. Skaug, and M. Brun (2022) Epistemic uncertainty quantification in deep learning classification by the delta method. Neural networks 145, p. 164–176. Cited by: §2.2.1. N. Radcliffe (2007) Using control groups to target on predicted lift: building and assessing uplift model. Direct Marketing Analytics Journal, p. 14–21 (English). Cited by: §A.3. [30] S. Raha, K. Khare, and R. K. Patra Computationally efficient laplace approximations for neural networks. In NeurIPS 2024 Workshop on Bayesian Decision-making and Uncertainty, Cited by: §2.2.1. C. Riquelme, G. Tucker, and J. Snoek (2018) Deep bayesian bandits showdown. In International conference on learning representations, Vol. 9. Cited by: §2.2.1. Y. Saito, S. Aihara, M. Matsutani, and Y. Narita (2020) Open bandit dataset and pipeline: towards realistic and reproducible off-policy evaluation. arXiv preprint arXiv:2008.07146. Cited by: §4.1. T. Schnabel, A. Swaminathan, A. Singh, N. Chandak, and T. Joachims (2016) Recommendations as treatments: debiasing learning and evaluation. In Proceedings of the 33rd International Conference on Machine Learning, Cited by: §1, §1, §2.2. U. Shalit, F. D. Johansson, and D. Sontag (2017) Estimating individual treatment effect: generalization bounds and algorithms. In Proceedings of the 34th International Conference on Machine Learning, Cited by: §1, §2.1. C. Shi, D. Blei, and V. Veitch (2019) Adapting neural networks for the estimation of treatment effects. In Advances in Neural Information Processing Systems, Cited by: §1, §2.1, §2.1, §3.2. Y. Su, X. Wang, E. Y. Le, L. Liu, Y. Li, H. Lu, B. Lipshitz, S. Badam, L. Heldt, S. Bi, et al. (2024) Long-term value of exploration: measurements, findings and algorithms. In Proceedings of the 17th ACM International Conference on Web Search and Data Mining, p. 636–644. Cited by: §2.2.1, §2.2. Z. Sun, H. Yang, D. Liu, Y. Weng, X. Tang, and X. He (2024) End-to-end cost-effective incentive recommendation under budget constraint with uplift modeling. In Proceedings of the 18th ACM Conference on Recommender Systems, p. 560–569. Cited by: §1.1, 4th item. A. Swaminathan and T. Joachims (2015) Batch learning from logged bandit feedback through counterfactual risk minimization. The Journal of Machine Learning Research 16 (1), p. 1731–1755. Cited by: §2.2. A. Swaminathan and T. Joachims (2017) Off-policy evaluation for slate recommendation. In Advances in Neural Information Processing Systems, Cited by: §1. M. J. van der Laan and D. Rubin (2006) Targeted maximum likelihood learning. The International Journal of Biostatistics 2 (1). External Links: Document Cited by: §3.2. A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin (2017) Attention is all you need. Advances in Neural Information Processing Systems. Cited by: §3.1. Y. Wang, D. Liang, L. Charlin, and D. M. Blei (2018) The deconfounded recommender: a causal inference approach to recommendation. Cited by: §1. C. Wei, B. Zelditch, J. Chen, A. A. S. T. Ribeiro, J. K. Tay, B. O. Elizondo, S. K. Selvaraj, A. Gupta, and L. B. D. Almeida (2024) Neural optimization with adaptive heuristics for intelligent marketing system. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, p. 5938–5949. Cited by: §1. R. Ying, R. He, K. Chen, P. Eksombatchai, W. L. Hamilton, and J. Leskovec (2018) Graph convolutional neural networks for web-scale recommender systems. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (KDD ’18), p. 974–983. Cited by: 5th item. W. Zhang, D. Zhou, L. Li, and Q. Gu (2020) Neural thompson sampling. arXiv preprint arXiv:2010.00827. Cited by: §2.2.1.