A STUDY OF APPROXIMATIONS AND DATA REDUCTION TECHNIQUES FOR KERNEL REGULARIZED LEAST SQUARES
Open Access
Author:
Straub, Benjamin Marshall
Graduate Program:
Statistics
Degree:
Master of Science
Document Type:
Master Thesis
Date of Defense:
April 02, 2018
Committee Members:
Bharath Kumar Sriperumbudur, Thesis Advisor/Co-Advisor Donald Richards, Committee Member David Russell Hunter, Committee Member Benjamin Adam Shaby, Committee Member
Keywords:
Kernel Methods Machine-Learning Ridge Regularization Random Projections Clustering Least-Squares
Abstract:
Researchers using machine-learning algorithms are encountering problems of data storage and computation time with the advent of Big Data in almost all aspects of life and work. Popular machine-learning techniques that perform computation- ally well on small datasets often suffer from computational efficiency problems on larger datasets. A second concern is the need for applied researchers to understand and interpret underlying phenomena, which can prove challenging with machine- learning algorithms. The kernel regularized least squares algorithm (KRLS) is a machine-learning technique that has great promise for allowing this understand- ing, but still suffers from computational issues. In this thesis, I study several data reductions techniques such as clustering, random sampling and random sketching within the KRLS framework in order to decrease computational time while seeking to only sacrifice a little in prediction performance.
