This research paper discusses the concept of predictive multiplicity, which refers to the presence of multiple competing models that offer nearly the same level of performance but yield conflicting results for individual data samples. This phenomenon poses significant challenges as it could lead to systematic exclusion, inexplicable discrimination, and unfairness in practical applications. The authors propose a novel framework that employs dropout techniques to explore these almost-equally-optimal models, also known as the Rashomon set. The proposed method has been found to outperform existing ones in terms of efficiency and speed in estimating predictive multiplicity metrics.
Publication date: 2 Feb 2024
Project Page: Not Provided
Paper: https://arxiv.org/pdf/2402.00728
