New Causal Bandit Algorithm Identifies Best Treatment with Guaranteed Correctness
Researchers use Track-and-Stop to confidently pick treatments while bypassing mediator effects.
This paper tackles the problem of identifying the best treatment in a causal bandit setting when the outcome is influenced by a mediating variable researchers want to exclude. The authors formalize the concept of natural direct potential outcome (NDPO) — the outcome under an intervention while blocking the path through the mediator. They first prove that NDPO can be identified from observable interventional distributions at the population level, laying the foundation for sequential decision-making under uncertainty.
To achieve sample-efficient identification with statistical guarantees, the team develops a fixed-confidence best-arm identification (BAI) algorithm based on the Track-and-Stop (TaS) framework. They solve the resulting semi-infinite optimization problem using a cutting-set method, ensuring δ-correctness (i.e., the probability of incorrect identification is bounded by δ) and asymptotic optimality in sample complexity. Empirical validation on the large-scale IPinYou advertising dataset demonstrates the algorithm's effectiveness in selecting treatments that maximize NDPO, offering a principled approach for applications in advertising, healthcare, and policy evaluation where mediator effects must be removed.
- Focuses on natural direct potential outcome (NDPO) to exclude mediator pathways from treatment evaluation.
- Uses Track-and-Stop with a cutting-set method to solve semi-infinite optimization, guaranteeing δ-correctness and asymptotic optimality.
- Validated on the large-scale IPinYou advertising dataset, showing practical viability for real-world causal bandit problems.
Why It Matters
Enables trustworthy treatment selection in advertising and healthcare when mediating variables distort outcome evaluation.