Off-campus WSU users: To download campus access dissertations, please use the following link to log into our proxy server with your WSU access ID and password, then click the "Off-campus Download" button below.
Non-WSU users: Please talk to your librarian about requesting this thesis through interlibrary loan.
Date of Award
Feature selection has been one of the hottest branch in machine learning and data
mining. In the past decade, high dimensional big data has proliferated in various areas,
such as text mining, bioinformatics, computer vision, e-commerce. The high dimensionality
brings great challenges for traditional statistical and machine learning techniques. Feature
selection is one effective technique handling the challenges. It can not only improve model
generalization performance and robustness, but also make model interpretable to better
understand the underlying mechanism of data generation.
In this thesis, we specifically study the multi-class classification problems using regularized multinomial logistic regression where the number of features is large and intrinsic
group structure of features exists. For multi-class classification, different classes might have
different related features and feature groups. Our goal is to improve mdoel performance by
incorporating the group information in the model and simultaneously achieving sparsity at
the feature and feature group level. Hence, we propose a class-conditional regularization
method for multinomial logistic regression for this purpose. For model optimization, an
efficient cyclic block coordinate descent algorithm is developed. Finally, we evaluate our
method on real-world datasets to demonstrate its effectiveness.
Li, Xiangrui, "Learning Sparse Features And Feature Groups For Multinomial Classification" (2020). Wayne State University Theses. 776.