A0370
Title: Skew-normal classification in high-dimensional data
Authors: Haesong Choi - Florida state university (United States) [presenting]
Qing Mai - Florida State University (United States)
Abstract: A considerable number of studies have been devoted to high-dimensional classification models under the assumption of normality. However, the normality assumption may be too restrictive in data analysis. Motivated by the data sets that exhibit asymmetry, including environmental, financial, and biomedical ones, we propose a high-dimensional discriminant analysis model called the SKNC model (short for SKew-Normal Classification). By incorporating the skew-normal distribution, the SKNC model is closely related to the LDA model but improves its flexibility on skewed data in classification. We further develop a novel classifier to estimate the SKNC model. To tackle the statistical challenge of the heavy-tailed distribution, we propose a robust estimation of parameters. We develop an efficient algorithm to adopt the penalized estimation. Theoretical results rigorously show that the SKNC model achieves variable selection and penalized estimation, especially in high-dimensional settings. We empirically demonstrate the superior performance of the SKNC model over existing methods in simulated and real datasets.
