Engineering Commons^™

Nonconvex Optimization For Statistical Learning With Structured Sparsity, Chengyu Ke

Operations Research and Engineering Management Theses and Dissertations

Sparse learning problems, known as feature selection problems or variable selection problems, are a popular branch in the field of statistical learning. When faced with a dataset with only a few observations but a large number of features, we are interested in extracting the most useful features automatically by solving an optimization problem. In this dissertation, we start by introducing a novel penalty function as well as an iterative reweighted algorithm to solve the group sparsity problem, a special type of feature selection problems. The penalty function, named group LOG, shows a better ability to recover the ground-truth compared to …