Causal Optimization Boosts LinkedIn Feed Marketing by 7.2%

Traditional large-scale targeting and recommendation systems often prioritize predictive scores, which are then fed into heuristic allocation rules. However, when the primary business objective is to maximize *incremental* impact—such as in marketing campaigns, incentives, or notifications—this predictive paradigm can lead to misallocation of resources, often targeting users who would have acted even without intervention. This research proposes a new decision-centric framework designed to optimize for causal effects under global constraints. The framework integrates three core components: a causal neural network, featuring a Transformer backbone, for estimating individual treatment effects; a Bayesian neural-bandit layer for uncertainty-aware exploration; and a dual-based large-scale linear-programming layer for constrained resource allocation. This integrated approach ensures that all components align with the single objective of maximizing incremental value. The framework also supports sequential contexts and multi-outcome, attribute-conditioned scoring through a Transformer encoder and outcome embeddings. Offline simulations and targeted architectural ablations were conducted, culminating in a successful online A/B test on LinkedIn Feed marketing traffic. This real-world deployment demonstrated a statistically significant 7.20% lift in the primary long-term-value metric, validating the feasibility and effectiveness of production-scale causal optimization under real business constraints. The authors also shared critical production lessons regarding causal training-data construction and cost/delivery control.