Offline RL frames next-best action as a constrained decision process
Original title: Next Best Action: Causal Constrained Offline RL for Business Action Targeting
In a rideshare coupon study, the authors report 19% higher lifetime value than the behavior policy.
The paper casts next-best-action selection as a constrained Markov decision process and adds off-policy causal correction inside its linear program. The resulting policy chooses actions from customer states while enforcing operational guardrails, and shadow prices describe constraint trade-offs. In a rideshare coupon case using recency, frequency, and monetary states, the authors report 19% higher customer lifetime value than the estimated behavior policy and 8% over a myopic baseline.
Unconfirmed: The lifetime-value gains are the authors' off-policy estimates, not a live deployment result.
- Evaluation uses model-based off-policy estimates from a temporal holdout.
- The demonstrated action space is binary coupon targeting under an observed budget.
- Shadow prices cover trade-offs such as fairness and user welfare.
- The authors describe the framework as action-agnostic for multiple interventions.
From the original article
1 min read at arxiv.orgNext Best Action (NBA) systems must select among customer interventions (e. g. , discounts, promotions, and lifecycle actions), maximize long-term economic value, satisfy complex operational constraints, and remain causally grounded.
By Surya Ramachandiran · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.