2024-03-29

Wei-Wen Hsu, Ph.D., Division of Biostatistics and Bioinformatics, Department of Environmental and Public Health Sciences, University of Cincinnati College of Medicine

"A Novel Exponential Loss Function for Multiclass Classification with Complex Imbalanced Data"

Multiclass classification has emerged as a significant focus across various domains including disease diagnosis, animal species classification, object identification, and face recognition. As the applications diversify and expand, there is a growing demand for effective approaches capable of handling multiclass classification within complex settings, particularly those characterized by imbalanced and high-dimensional longitudinal data structures. Traditional classifiers often struggle in scenarios marked by imbalanced class distributions, tending to misclassify minority class instances as belonging to majority classes. Consequently, predictive accuracy suffers significantly, especially for minority classes. This challenge is exacerbated in longitudinal and high-dimensional contexts, where the classification performance is expected to be further compromised. In this paper, we propose a novel two-stage classification framework tailored to address the complexities of multiclass imbalanced data. Initially, we employ techniques based on natural cubic splines to efficiently extract features from longitudinal data. Subsequently, we introduce a weight-adjusted margin-based exponential loss function, coupled with the group LASSO penalty, designed specifically for multiclass classification in high-dimensional settings. Notably, unlike existing methods, our proposed approach constitutes a single optimization procedure capable of simultaneous feature selection across all classes. We evaluate the empirical performance of our method through extensive simulations in finite sample sizes and further demonstrate its efficacy through real-world applications.

Stay connected TwitterFacebook LinkedIn YouTubeInstagram