Comprehensive Exam of CMSE Ian Loveless

April 22, 2025, 1pm, EB 3405B

 

By Ian Loveless

 

Abstract:

Machine learning models often assume a specific structure in data—categorical, ordinal, or continuous—without explicitly verifying whether these assumptions hold. This dissertation develops and validates a method for assessing ordinality in the feature space and assesses how ordinality impacts model performance. The objective is to determine when structure-aware models outperform conventional classification approaches. Two primary applications guide this research: (1) ordinal classification in medical imaging, particularly for pneumoconiosis and prostate cancer grading, and (2) structural characterization in biological systems, specifically pancreatic cancer subtyping using single-cell and spatial transcriptomics data. By unifying these domains, this work proposes novel pairwise separability measures and extends ordinal detection methods to high-dimensional feature spaces. Findings demonstrate that quantifying structure in data can inform optimal model selection but also optimal class

orderings, ensuring that machine learning models align with the underlying data structure. The results have implications for medical diagnostics, deep learning model interpretability, and cancer biology.

 

 

Committee:

Adam Alessio (chair)

Guanqun Cao

Jianrong Wang

Bin Chen