Friday, December 13, 2019, 1 pm to 3 pm
304 Russ Engineering
Ph.D. Committee: Drs. Guozhu Dong (advisor), Keke Chen, Krishnaprasad Thirunarayan, Pratik Parikh (BIE), and Hemant Purohit (George Mason University)
Many classification algorithms have been proposed over the last 50 years. However, each classification algorithm may perform poorly for certain applications (or data sets), and it is of interest to understand why and how such poor performance happens. Poor performance can occur due to the data being especially hard to classify, or because of inherent weaknesses in a type of classifier.
In this dissertation we answer these questions by examining the richness and simplicity of easily correctable opportunities that are present when given classification algorithms are applied. We accomplish these by using a novel, pattern-based approach to systematically identify weaknesses of different classifiers, and to evaluate the potential of improvement on classification performance. Moreover, we develop a number of useful metrics to consistently measure and compare the richness and correctability of multiple significant opportunities. It is hoped that the insights and techniques developed here can provide guidance to future classification researchers, and to machine learning practitioners concerning classifier weaknesses, and can help them to better understand what classifiers are best suited to given difficult-to-classify applications.
Log in to submit a correction for this event (subject to moderation).