From Prediction to Incrementality: Causal Optimization for Large-Scale Targeting and Recommendation

The paper presents a decision-centric framework for large-scale targeting and recommendation systems that optimizes causal effects under global constraints, rather than the traditional paradigm of optimizing predictive scores under heuristic allocation. The authors state: 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. The framework aligns 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 causal model is a modified version of DragonNet, which jointly estimates mu1(X), mu0(X), and e(X) in a single forward pass. The authors extend DragonNet with three key enhancements: "First, we feed sequential user-event features through a transformer layer 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, forming a DragonBandit policy that explores over incremental action values rather than predicted responses."

For exploration, the paper uses Neural Thompson sampling via Laplace approximation, which approximates the posterior over network parameters by a Gaussian centered at the MAP solution via the linearized Laplace approximation (LLA). This allows the same supervised checkpoint can be deployed with or without exploration and no retraining is required. The authors note they also evaluated Bayes by Backprop offline, but use LLA in production because it adds uncertainty sampling to the existing checkpoint with substantially lower training and serving overhead.

For allocation, the paper formulates the problem as a large-scale linear program with constraints including guardrail constraints, frequency caps, and binary decision variables. To scale, they use a smoothed dual-decomposition method where the resulting dual is differentiable with Lipschitz gradient and the primal minimizer decomposes per user, giving per-iteration cost linear in U·I versus O((UI) cubed.5) for interior-point methods. They also use warm-starting: "we warm-start the dual from the previous period's 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."

The Transformer architecture uses a touchpoint sequence with temporal position encoding combining a days-since embedding, day-of-week embedding, and the small-embedding-friendly sinusoidal positional encoding. Multi-head self-attention is applied, and pooling by taking only the [CLS] output as the shared representation. For multi-outcome support, the model uses outcome embeddings where each outcome tower drops the final sigmoid and returns hidden logits representing the member's expected state under T=1 and T=0, and a single shared outcome embedding matrix encodes the K products. The outcome embeddings are regularized through an autoencoder: The autoencoder is trained jointly with the main task using MSE reconstruction, which serves as semantic regularization. The final outcome logits are L2-normalized dot products with learnable log-inverse-temperatures.

The complete loss function is: L = L YT=1 + L YT=0 + L T + L tarreg + λ recon L recon, where L tarreg is based on the efficient influence function (EIF) for the ATE, used as a training regularizer, not as a substitute for identification assumptions.

Offline simulations use the Open Bandit Dataset, with "the random-policy subset, synthetically mapping the original 34 products that were recommended to >400K users to 5 distinct actions." The authors compare five approaches: Incremental Modeling + Constrained Optimization, Bandit Incremental Modeling + Constrained Optimization, Propensity Modeling + Constrained Optimization, Incremental Modeling + Ranking, and Propensity Modeling + Ranking. In single-turn evaluation, incremental scores lead to higher rewards compared to propensity scores, and incremental targeting leads to a higher percentage of no-recommendations as the engine is able to identify members likely to convert organically. The results show Incremental Modeling coupled with Constrained Optimization provides the highest overall performance with average net return of 0.46 ± 0.13.

In multi-turn evaluation with biased training data, 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, and the benefit of exploration is more pronounced when the initial training-data bias is larger.

The ablation study tested 8 configurations and found that 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, consistent with the dense features being prognostic rather than treatment-effect modifiers.

For production deployment, the paper describes training-data construction: The context must contain only information available before treatment, while treatment and outcome are assigned from subsequent non-overlapping windows. They use a randomized date D with W D = 90 days, W T = 7 days, and W C = 30 days, setting T=1 for qualifying interactions in [D, D+7) and Y=1 for conversions in [D+7, R]. They also describe cost and delivery control, including a feedback controller that adjusts the round-level cost target based on the spend-realization ratio and a pacing curve. The authors note: "An early A/B test run without the above controls showed the treatment arm under-delivering... A raw total-bookings comparison then penalizes the treatment for delivering less rather than for choosing worse, confounding policy quality with delivery volume."

The serving architecture is a batch pipeline where "DragonBandit scores μ1, μ0, and e for the eligible population and samples incremental objectives via last-layer linearized Laplace Thompson sampling; a dual-decomposition solver computes the constrained allocation under global guardrails, warm-started from the previous period's dual λ*; and the feedback controller adjusts budget caps from realized deliverability before assignments are materialized."

For the online experiment, the authors used an eight-week online A/B test on LinkedIn Feed marketing traffic, against the deployed business-as-usual (BAU) targeting stack. They 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. The results: the treatment arm achieved a statistically significant +7.20% lift (p = 0.041, 95% CI: [0.31%, 14.09%])" in the primary long-term-value metric.

The paper's contributions are: "(i) a decision-centric formulation that aligns causal estimation, exploration, and constrained allocation under a single objective; (ii) 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 (iii) a productionized pipeline combining the causal head with neural-bandit exploration and a dual-based large-scale LP solver."

Based on the paper, here are the specific improvements you can make to AI systems and what the improved systems can do:

1. Decision-centric objective alignment

• Replace predictive-score optimization with a unified objective that directly maximizes causal incremental impact (e.g., ATE or CATE) under global resource constraints.

• The improved system will allocate interventions (e.g., discounts, notifications, ads) to users who would not act otherwise, reducing wasted spend on sure-thing users and increasing net return per unit of budget.

2. Causal neural network with temporal transformer encoder

• Feed sequential user-event features (recency, engagement cadence) through a transformer layer before treatment-effect estimation, using temporal position encodings (days-since, day-of-week, sinusoidal).

• The improved system will capture time-varying user state and interaction patterns, leading to more accurate individual treatment-effect estimates (μ1, μ0) and propensity scores (e), especially for users with long or irregular touchpoint histories.

3. Multi-outcome embedding head for cold-start products

• Use a shared outcome-embedding matrix and L2-normalized dot-product logits with learnable inverse-temperatures, regularized by an autoencoder (MSE reconstruction).

• The improved system can score newly introduced products or actions without retraining a new output head, enabling rapid scaling to new inventory or campaigns while preserving outcome AUROC and uplift AUUC.

4. Uncertainty-aware exploration via Laplace-approximated neural Thompson sampling

• Add last-layer linearized Laplace approximation (LLA) to the causal predictor, sampling incremental action values (not predicted responses) from the posterior.

• The improved system will explore actions with high uncertainty in incremental impact, overcoming biased training data (e.g., from historical targeting policies) and improving long-term performance in multi-turn settings, especially when initial data bias is large.

5. Large-scale constrained allocation via smoothed dual decomposition

• Formulate allocation as a linear program with guardrails, frequency caps, and binary decisions; solve via a differentiable dual with per-user primal decomposition and warm-starting from previous λ*.

• The improved system will handle millions of users and items with per-iteration cost linear in U·I (vs. cubic for interior-point), meet global constraints in real time, and degrade gracefully (99% of optimum) under stable input distributions, with SLA fallback.

6. Feedback-controlled budget pacing

• Integrate a feedback controller that adjusts round-level cost targets based on spend-realization ratio and a pacing curve.

• The improved system will avoid under-delivery or over-delivery, ensuring the treatment arm receives the intended volume of interventions, preventing confounded comparisons in A/B tests and real-world deployment.

7. Training-data construction with non-overlapping windows

• Use a randomized date D, with treatment window W T and outcome window W C strictly after W T, and context features only from before D.

• The improved system will produce unbiased causal estimates by eliminating leakage between treatment assignment and outcome measurement, making it safe for marketing, incentives, and notification systems.

8. Agentic campaign-setup tool for clean online experimentation

• Build a tool that duplicates campaigns into budget-split variants, freezes static audiences, audits segment lineage (no control depends on treatment-derived decisions), and enables symmetric rollback.

• The improved system will run statistically valid online experiments with minimal operational overhead, isolating policy quality from delivery volume, and enabling reliable lift measurement (e.g., +7.20% LTV lift with p=0.041).

9. Single-checkpoint deployability with optional exploration

• Use the same supervised checkpoint for both exploitation and exploration (via LLA), without retraining.

• The improved system will switch between deterministic and exploratory modes at serving time, reducing training/serving overhead while retaining the ability to adapt to non-stationary environments.

10. Semantic regularization for outcome embeddings

• Jointly train the autoencoder with the main task to enforce semantic structure in outcome embeddings.

• The improved system will generalize better to unseen product-action combinations and maintain stable uplift estimates across heterogeneous item catalogs, as shown by the ablation study where outcome embeddings preserved AUROC and improved uplift AUUC.

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.