Predictive Disparity in Modern Machine Learning Applications - Sources and Impact

Explore a thought-provoking conference talk from the Big Data Conference 2024 where University of Washington's Jamie Morgenstern delves into the critical examination of predictive disparities in modern machine learning systems. Learn how statistical models deployed in influential domains like healthcare, political campaigns, and financial instruments have historically shown varying levels of predictive accuracy across different demographic groups. Understand the key factors contributing to these disparities, including limited model expressiveness, data availability constraints for marginalized populations, measurement inconsistencies, and varying feature relevance across populations. Gain insights into how these challenges manifest in contemporary machine learning applications, particularly in facial recognition systems and generative AI, and discover the evolving nature of performance differences across diverse populations in modern ML implementations.