Robust and Scalable Representation Learning for Massive Graphs and Time Series Data

The rapid growth of complex, high-dimensional data—such as graphs and time series—has created a pressing need for robust and scalable representation learning methods. Geometric machine learning offers promising tools for analyzing non-Euclidean data structures, yet it remains limited by challenges in scalability, interpretability, and performance across heterogeneous data sources. This dissertation addresses these challenges by proposing novel, theoretically grounded, and computationally efficient methods that advance representation learning for unstructured, multi-modal, and streaming data. Through a unified framework, this work enhances the interpretability, robustness, and scalability of machine learning models applied to real-world problems in fields such as bioinformatics and cyber-physical systems.