Uplifting conversion across the acquisition funnel with personalization using contextual bandits on AWS
aws ·
This article details how Amazon Payments implemented a multi-objective contextual bandit system on AWS to optimize personalization across a three-stage product acquisition funnel. By utilizing Linear UCB (LinUCB) within Amazon SageMaker AI, the team conditioned content selection on individual customer behavioral signals rather than static segments. The model simultaneously optimized for application starts, submissions, and approvals using a weighted combination of scores to address the 'seesaw problem' where optimizing one stage might negatively impact another. Deployment relied on a weekly batch processing architecture that updated models based on delayed feedback, storing recommendations in a low-latency key-value store for serving. Results showed a high single-digit percentage lift in final-funnel conversion for one customer segment, while another saw no improvement due to insufficient content quality rather than algorithmic failure. The authors emphasize that while generative AI expands the pool of creative variations, contextual bandits are essential for selecting the most effective variant for each visitor. Key lessons include the importance of validating content pools through thorough exploration and using offline evaluation alongside A/B testing to ensure robust performance.
aws