Next Best Action: Causal Constrained Offline RL for Business Action Targeting

Key points

• We address this gap by formulating NBA as a Constrained Markov Decision Process (CMDP).1
• We then integrate off-policy causal correction of the estimated customer dynamics directly into the Linear Programming (LP) formulation of the CMDP.1
• We validate our proposed offline Reinforcement Learning (RL) framework on an industry case study involving rideshare coupon targeting, where customer states are defined by Recency, Frequency, and Monetary (RFM) stages.1
• We evaluate performance using model-based off-policy evaluation, with causal dynamics estimated from a temporal holdout.1

Sources (1)

• [1]Next Best Action: Causal Constrained Offline RL for Business Action Targeting
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 19, 05:51 PM

  Next 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.

  We address this gap by formulating NBA as a Constrained Markov Decision Process (CMDP).

Extractive summary: sentences quoted from the sources.