Differentially Private Algorithms for Statistical Estimation Problems

Watch a 52-minute conference talk from the Harvard CMSA Big Data Conference where Columbia professor Rachel Cummings explores recent developments in differential privacy (DP) algorithms across three key statistical domains. Dive into how Thompson sampling inherently satisfies DP requirements in online learning with bandit feedback, discover the first DP algorithms for synthetic control in causal inference, and learn why traditional approaches to imbalanced learning fail when combined with DP algorithms. Understand how differential privacy provides rigorous bounds on information leakage while maintaining analytical accuracy through calibrated randomness injection. Explore novel solutions for privately generating synthetic minority points in datasets with underrepresented classes, based on collaborative research with Marco Avella Medina, Vishal Misra, Yuliia Lut, Tingting Ou, Saeyoung Rho, and Ethan Turok.