Data Science Talk Series 

Trustworthy AI from Hospital Beds to Global Marketplaces

Abstract: As artificial intelligence increasingly drives critical decisions in healthcare and e-commerce, ensuring model trustworthiness through fairness, interpretability, and privacy has become paramount. This talk presents novel methodological advances in survival analysis - a statistical framework for predicting time-to-event outcomes - and demonstrates their practical applications from healthcare to Amazon's marketplace optimization. In healthcare, we introduce innovative pseudo-value-based deep learning frameworks that address fundamental challenges in survival analysis: censoring (incomplete observations), competing risks (multiple possible outcomes), and multi-state transitions (complex disease progressions). To ensure fairness, we propose novel constraints that significantly improve predictions for underrepresented groups while maintaining high accuracy. For multi-institution collaboration scenarios, we develop privacy-preserving federated frameworks (FedPseudo, Fedora, FedPRF, FairFSA) that enable distributed learning without raw data sharing, demonstrating up to 17% improvement over baselines. This talk then showcases how these survival analysis principles extend to causal inference problems at Amazon, addressing severe data imbalances, cold-start problems, and sparse outcomes in optimizing fulfillment strategies across billions of marketplace offers.