Loading...
Robust machine learning for complex data: enhancing efficiency and powering healthcare solutions
Li, Xiangyu
Li, Xiangyu
Citations
Altmetric:
Advisor
Editor
Date
Date Issued
2024
Date Submitted
Keywords
Collections
Research Projects
Organizational Units
Journal Issue
Embargo Expires
Abstract
In today's data-driven world, the demand for extracting meaningful insights from vast datasets is increasing alongside the growing complexity of machine learning models. Traditional machine learning faces constraints, including development in noiseless or low-noise environments and heavy reliance on assumptions that may not hold in real-world scenarios. A notable limitation is their dependence on Euclidean distances, making them vulnerable to perturbations introduced by outliers or inherent noise in real-world datasets. To address these challenges, this research focuses on designing machine learning algorithms that exhibit robustness and adaptability to large-scale, heterogeneous data. Key contributions include adaptive objectives for dimensionality reduction, a novel method for robust graphical representation, and structure-awareness in representation learning. These advancements bridge the gap between complex data structures and actionable insights in both medical and computational domains.
Effective optimization is crucial in data-driven research, especially as many machine learning methods suffer from non-convex or non-smooth optimization problems. My research explores the Proximal Alternating Linearized Minimization (PALM) technique for robust convergence and superior prediction accuracy. Additionally, the Generalized Iteratively Reweighted Method (GIRM) is found to effectively manage non-convex objectives. Furthermore, my research introduces Riemannian Optimization, which transforms simplex problems into a smooth manifold, and proposes a Tangent Riemannian Gradient Descent (T-RGD) method that improves the efficiency and convergence in optimization. These contributions provide robust and efficient tools for reliable analysis across diverse real-world datasets.
As the complexity and volume of medical data increase, advanced AI-driven methodologies become necessary. This research leverages our proposed machine learning methods for enriched representation learning in longitudinal chest X-ray analysis, enabling early diagnosis. Additionally, our research delves into bioinformatics, uncovering protein interactions and exploring drug repurposing for COVID-19. These machine learning-driven approaches contribute to improved healthcare outcomes and deeper scientific insights.
In conclusion, this thesis advances machine learning's theoretical foundations and practical applications, providing robust models, streamlined optimization methods, and AI-powered healthcare solutions. It addresses real-world complexities for improved healthcare and scientific insights.
Associated Publications
Rights
Copyright of the original work is retained by the author.