Date of Award
Campus Access Dissertation
Doctor of Philosophy (PhD)
Machine learning is a rapidly evolving field with vast potential. At its core, machine learning research aims to develop broadly applicable algorithms capable of improving their performance independently and without explicit programming. In recent years, we've seen fantastic progress in autonomous driving, speech recognition, and recommendation systems. Such success can be attributed to powerful learners and a good representation space crafted specifically to ease a subsequent learning task. Good representation learning approaches preserve as much information about the input data as possible and incorporate general information about the world around us that is not task-specific but would likely be beneficial for a learner to solve the task at hand. The extensive adoption of ML in different domains affirmed several well-known issues and discovered new challenges requiring special attention from machine learning and domain application experts. This dissertation focuses on a topic of space representation and explores three challenges. The first challenge is the clustering of sparse high-dimensional data, where we must learn a good representation for efficient and useful clustering. Second, multi-modal data fusion focuses on composing a joint representation space based on a set of unimodal representations learned from multiple modalities. Third, machine learning fairness is integral to a good representation and has recently become an important topic in the ML community. All of our methods have been validated on real-world data and are shown to outperform current baseline approaches.
Andreeva, Olga, "A Study of Several Critical Machine Learning Topics and Their Applications in Health Informatics" (2022). Graduate Doctoral Dissertations. 762.