Paper co-authored by Moritz Hardt receives 10-Year Test of Time Award at ACM STOC 2025

Prague – The paper "Preserving Statistical Validity in Adaptive Data Analysis", published at ACM STOC 2015, was honored with a 10-year Test of Time Award at the 57th ACM Symposium on Theory of Computing (STOC 2025), which took place in Prague in June. ACM STOC is the premier publishing venue in theoretical computer science. Moritz Hardt, who is the Director of the Social Foundations of Computation Department at the Max Planck Institute for Intelligent Systems, shares the award with his co-authors Cynthia Dwork, Vitaly Feldman, Toniann Pitassi, Omer Reingold, and Aaron Roth.

Traditional theory in machine learning and statistics assumes that a data analyst decides on what to do up front, then collects data and runs the analysis on the collected data. When working with data, however, practitioners typically work adaptively, deciding on new steps after seeing the results of previous ones. Evaluations, tests, and analyses therefore depend on prior interactions with the data. This feedback loop between method and data is what characterizes adaptive data analysis.

The 10-year Test of Time Award recognizes the work that founded the area of adaptive data analysis. The authors introduced a formal model of adaptive data analysis and observed that adaptivity invalidates almost all classical statistical theory. The main result is a new meta theorem: Any differentially private method is also statistically valid in the adaptive analyst model. This new connection between privacy and statistics provided the first theoretical results on adaptive data analysis and sparked much follow-up work.

For additional reading, see:

The reusable holdout: Preserving validity in adaptive data analysis (Google Blog, 2015)

Adaptive data analysis (Blog post 2015)

Guilt-free data reuse (CACM, 2017)