Date of Award
Summer 7-14-2026
Document Type
Dissertation
Publication Status
Version of Record
Submission Date
August 2026
Department
Computer and Electrical Engineering and Computer Science
College Granting Degree
College of Engineering and Computer Science
Department Granting Degree
Electrical Engineering and Computer Science
Degree Name
Doctor of Philosophy (PhD)
Thesis/Dissertation Advisor [Chair]
Yufei Tang
Abstract
Spatiotemporal and structured data are widely observed in environmental monitoring, sensor networks, ocean science, transportation, and graph-based applications. In these systems, observations are rarely independent; instead, prediction and interpretation often depend on relationships among variables, locations, time steps, entities, and labels. However, existing models are commonly studied from an architecture-centered perspective, which provides limited insight into how relational information is obtained, represented, utilized, and explained. This dissertation develops a relationship-aware perspective for model learning and explaining in spatiotemporal and structured systems. It first introduces a relationship-aware taxonomy that characterizes predictive models according to predefined relationships, data-inferred relationships, and implicitly captured dependencies. It further provides data-side, task-side, and model-side guidance for identifying relational evidence, selecting suitable modeling strategies, and classifying existing methods. Based on this foundation, the dissertation studies three representative problems. First, it develops a missing-aware latent relationship learning framework for robust multivariate time-series forecasting under severe sensor missingness, where imputation, dependency learning, and spectral–temporal forecasting are jointly optimized. Second, it presents a sparse spatiotemporal relationship modeling framework for ocean salinity imputation from irregular drifter trajectories, using global dependency modeling and diffusion-adversarial learning to reconstruct missing salinity values. Third, it extends relationship-aware modeling toward explainability by developing a self-explainable framework for multilabel graph learning, where predictions and label-specific relational explanations are learned jointly. Together, these studies demonstrate that relationships provide a unifying lens for designing robust, accurate, and interpretable models across different tasks and data conditions. This dissertation contributes methodological advances and a coherent conceptual and practical framework for relationship-aware learning and explaining in complex real-world systems.
Recommended Citation
Feng, Yingqi, "RELATIONSHIP-AWARE MODEL LEARNING AND EXPLAINING FOR SPATIOTEMPORAL SYSTEMS" (2026). Electronic Theses and Dissertations. 405.
https://digitalcommons.fau.edu/etd_general/405