05/24/2023
By Parisa Hajibabaee
Ph.D. in Computer Science Candidate: Parisa Hajibabaee
Defense Date: Wednesday, June 07, 2023
Time: 10 to 11:30 a.m. EST
Location: This will be a virtual defense via Zoom
Committee Members:
Reza Ahmadzadeh (co-advisor), Ph.D., Assistant Professor, Computer Science, UMass Lowell
Farhad Pourkamali-Anaraki (co-advisor), Ph.D., Assistant Professor, Department of Mathematical and Statistical Sciences, University of Colorado Denver
Scott Stapleton, Ph.D., Associate Professor, Mechanical Engineering, UMass Lowell
Mohammad Amin Hariri-Ardebili, Ph.D., Affiliated Research Associate, Engineering and Applied Science, University of Colorado Boulder
Abstract:
Recent years have witnessed an explosion in the volume and complexity of data, presenting significant challenges for researchers and practitioners across a wide range of fields. In response, machine learning algorithms have become increasingly important tools for analyzing and making sense of this data. The main focus of this dissertation is centered around the intersection of machine learning and engineering applications within three projects:
The first project investigates how to improve the landmark selection technique to produce accurate low-rank approximations of kernel matrices on high-dimensional imbalanced data sets. The proposed approach aims to find a concise set of exemplars or landmarks to reduce the number of similarity measure evaluations, while also regulating tradeoffs between the quality of landmarks and resource consumption.
The second project addresses the challenge of effectively analyzing high-dimensional scientific data with class imbalance by combining dimensionality reduction (DR) techniques and oversampling. DR techniques are crucial for feature extraction and data visualization in scientific and engineering applications. The proposed framework systematically investigates the integration of different DR techniques and oversampling for analyzing scientific data.
The third project focuses on improving uncertainty quantification in machine learning models, specifically in the domain of regression problems with high-dimensional, skew heavy-tailed data. The proposed non-conformity measure based on data-dependent weights enhances the efficiency and conditional coverage of prediction intervals in conformal prediction.
These research findings contribute to the advancement of the field and offer insights that can inform future research and practice in machine learning and engineering applications.