A0173
Title: Model-free selective inference: Selecting trusted decisions from black boxes
Authors: Ying Jin - Harvard University (United States) [presenting]
Emmanuel Candes - Stanford (United States)
Abstract: Many decision-making and scientific discovery processes aim to identify candidates whose unknown outcomes satisfy a desired property, e.g., drugs with high binding affinities to a disease target in drug discovery. While predictive AI excels in accelerating such processes, ensuring the reliability of AI remains a challenge in critical situations. In the example of drug discovery, any false lead in shortlisted drug candidates may incur substantial costs in later stages, which harms its initial promise. A model-free selective inference framework is introduced to identify candidates whose unobserved outcomes exceed user-specified values with the assistance of any prediction model. The framework controls the average proportion of false positives (FDR) among the selected set of units without any modelling assumptions on the data distribution. In addition, new ideas on dealing with distribution shifts are discussed between training and new samples, a scenario often encountered in applications. Via several empirical studies in drug discovery, we show that the methods help scientists narrow down the drug candidates to a manageable size of promising ones with finite-sample error control.
