Graduation Semester and Year
Summer 2024
Language
English
Document Type
Dissertation
Degree Name
Doctor of Philosophy in Computer Science
Department
Computer Science and Engineering
First Advisor
Won Hwa Kim
Second Advisor
Gautam Das
Third Advisor
Guorong Wu
Fourth Advisor
Chengkai Li
Fifth Advisor
Dajiang Zhu
Abstract
Deep learning has profoundly transformed machine learning by offering sophisticated data representations, yet effectively incorporating structural information remains a challenge. Structural data, whether explicit or implicit, has the potential to significantly enhance the performance of deep learning tasks. This research investigates the benefits of structural information across three crucial tasks: classification, clustering, and segmentation. For explicit structural data, where inputs are directly represented as graphs, we investigate graph-level classification in brain connectivity networks. We introduce the Multi-resolution Edge Network (MENET), a novel framework designed to identify disease-specific connectomic benchmarks with high discriminatory power across diagnostic categories. MENET leverages graph-level representations to capture intricate patterns in brain networks, providing crucial insights into disease progression. Moreover, we extend this concept to broader brain connectivity analysis by enriching statistical inferences via latent space graph embedding. This enables the differentiation between preclinical and prodromal stages of Alzheimer's Disease, revealing significant regions of interest that distinguish affected areas. For multi-group analyses, our approach improves differentiation between stages like Early Mild Cognitive Impairment (EMCI) and Late Mild Cognitive Impairment (LMCI), while incorporating graph regularization to enhance stability and differentiation. In cases involving implicit structural data, where relationships among data samples are not immediately apparent, we first propose a deep contrastive clustering method utilizing k-nearest neighbor graphs derived from the dataset. This label-free approach effectively groups data samples without predefined labels or annotations. Building on this clustering framework, we extend its application to semantic segmentation without annotations. By leveraging the clustering results and integrating them with a foundational segmentation model, we accurately delineate coherent regions within complex datasets. This combination enables the production of meaningful semantic segmentation, revealing intricate patterns and boundaries. Comprehensive experimental evaluations across diverse datasets emphasize the pivotal role structural information plays in enriching deep learning frameworks. By optimizing latent space representations tailored to unsupervised tasks, structural data significantly improves clustering, classification, and segmentation. This research underscores the immense potential of structural information to refine and enhance deep learning methodologies, offering valuable contributions to these widely studied fields while establishing a robust framework for future research.
Keywords
Deep learning, Structural information, Graph-level classification, Contrastive clustering, Semantic segmentation, Latent space embedding, Implicit and explicit structural data
Disciplines
Artificial Intelligence and Robotics | Data Science | Other Computer Sciences | Statistical Methodology | Theory and Algorithms
License
This work is licensed under a Creative Commons Attribution 4.0 International License.
Recommended Citation
Ma, Xin, "Advancing Deep Learning with Graph-Based Structural Insights: From Graph Classification to Semantic Segmentation" (2024). Computer Science and Engineering Dissertations. 259.
https://mavmatrix.uta.edu/cse_dissertations/259
Included in
Artificial Intelligence and Robotics Commons, Data Science Commons, Other Computer Sciences Commons, Statistical Methodology Commons, Theory and Algorithms Commons