Date of Award
Doctor of Philosophy
Industrial and Manufacturing Systems Engineering
Industrialand Manufacturing Systems Engineering
Availability of data with hundreds of variables in the current era, emphasizes on the impor- tant of feature selection methods. Feature selection methods reduces computation time, improves prediction performance, and helps a better understanding of the data in the machine learning applications.
The main topic of this dissertation is developing new feature selection algorithms for data sets with continuous or binary response with consideration of feature interaction. Based on the literature, discovering interaction effect is a notoriously challenging task. In the first paper, we proposed a first optimization approach that can detect interacting features and the exact combination of these variables. The method was applied on a simulation experiment of the soybean data set and showed a %95 chance of correctly detecting second-order to fifth-order interaction effects between genes. The second paper discussed the importance of considering interacting features in feature selection. We developed a novel feature selection method for data sets with continuous response variables and established a computational framework for selecting a subset of features with significant main effects and a subset of features with significant interaction effects. And finally in the third paper, we expand our method for data sets with binary response and proposed a feature selection model that classifies two finite point sets in n-dimensional features space with consideration of feature interaction. Methods were applied on on the publicly available real-world databases and showed improvements in prediction analysis.
Maryam Nikouei Mehr
Nikouei Mehr, Maryam, "Discovering interacting features for prediction of response" (2020). Graduate Theses and Dissertations. 18053.
Available for download on Thursday, June 16, 2022