A0863
Title: ML-powered outlier detection: False discovery rate control and derandomization
Authors: Yaniv Romano - Technion---Israel Institute of Technology (Israel) [presenting]
Abstract: The focus is on recent advancements in outlier (or out-of-distribution) detection, highlighting how conformal inference plays a pivotal role in creating outlier detection algorithms that control the false discovery rate. After outlining the advantages of using conformal p-values for this task, an inherent limitation of this approach is addressed: its randomized nature. Such randomness often leads to different outcomes when analyzing the same test data, complicating the interpretation of findings. To alleviate this issue, a principled solution is presented to make conformal inferences more stable by leveraging suitable conformal e-values instead of p-values to quantify statistical significance. The landscape of machine learning and multiple hypothesis testing is navigated to ensure that conclusions extracted from any complex outlier detection model are reliable, stable, and reproducible.
