Analysis of Supersaturated Designs
Supersaturated designs have been widely studied in the literature. In a supersaturated design, the number of variables exceeds the number of runs. Most of existing studies focus on the construction of such designs, and the analysis methods have not been sufficiently explored. In this project, the researchers investigate an application in credit card industry. Credit card fraud is one of the biggest threats to business establishments today. Superior fraud detection is based on analyzing the abundance of past transactional and case disposition data that goes into building and finely tuning neural network models and other advanced analytics. This project tackles the feature selection as one of the bottlenecks in the model building. The focus is to find out the best subset of variables such that a neural network model built with this variable set will have an optimal detection performance. The researchers explore several variable selection methods including LASSO, the Elastic Net, and the Dantzig Selector, and they compare their performances based on prediction accuracies. They also propose a new method called the high dimensional stepwise selection method and apply it to the case study. Most of these methods are computationally intense, so the support of MSI is very important to the project.
A bibliography of this group’s publications is attached.
Return to this PI's main page.