A1415
Title: On discriminant analysis for the statistical few-shot learning of skewed data
Authors: Yana Melnykov - University of Alabama (United States) [presenting]
Semhar Michael - South Dakota State University (United States)
Andrew Simpson - South Dakota State University (United States)
Abstract: Few-shot learning with many classes and limited observations per class poses challenges, especially in high-dimensional settings. Traditional classification methods like linear or quadratic discriminant analysis assume that each class follows a multivariate normal distribution. However, quadratic discriminant analysis may suffer from unstable covariance estimates due to insufficient class-specific data. A recently developed approach based on clustering covariance matrices allows more flexible information sharing across classes. To relax the assumption of normality, a transformation-based procedure capable of handling skewed data is proposed. Monte Carlo simulations show promising results in parameter estimation as well as the classification accuracy.
