Causal Optimization Framework for Large-Scale Targeting and Recommendation

A novel framework focused on decision-making for extensive targeting and recommendation systems has been introduced, seeking to enhance causal impacts within global constraints instead of depending on predictive scores. This framework comprises three key elements: a causal neural network featuring a Transformer backbone for estimating individual treatment effects, a Bayesian neural-bandit layer that facilitates uncertainty-aware exploration, and a dual-based large-scale linear programming layer for constrained distribution. Additionally, it accommodates sequential context and multi-outcome, attribute-conditioned scoring via a Transformer encoder and outcome embeddings. This method tackles the issue of resource misallocation in marketing efforts, incentives, and notifications, where predictive scores frequently target users who would have engaged regardless. The framework underwent evaluation through offline simulations on a public bandit dataset. The paper can be accessed on arXiv with the identifier 2608.10182.