Online Program

Permutation-Based Uncertainty Quantification (308353)

*Vaidehi Ulhas Dixit, North Carolina State University  
Ryan Martin, North Carolina State University 

Keywords: Cesaro average, lasso, mixture models, permutations, recursive estimators, uncertainty quantification

Machine learning focuses largely on parameter estimation and prediction, but uncertainty quantification is important as well, especially in scientific applications. Motivated by a case where the estimator inherently depends on the data ordering, we propose to quantify uncertainty via a permutation distribution obtained by re-evaluating the estimator over multiple data permutations. For situations where the estimator does not inherently depend on the data ordering, the aforementioned method can still be employed after suitably manufacturing order-dependence.