Abstract
Connected systems generate large volumes of time series data that pose fundamental challenges for predictive analytics. These systems are inherently distributed, often operating under strict privacy constraints that preclude centralized data aggregation while still requiring collaborative modeling across heterogeneous units. At the same time, many predictive tasks rely on rich auxiliary information that is available during model development but absent at deployment, raising challenges in achieving robust generalization without inducing reliance on such contextual signals. Moreover, real-world connected systems exhibit dynamically evolving structures, where both the number of observed sequences and their organizational patterns vary over time, demanding models that can flexibly adapt to structural changes without retraining.
This dissertation investigates how predictive models for connected systems can remain accurate, privacy-preserving, and robust under distributed data ownership, and dynamic structural variation. To address this overarching question, this dissertation develops three methodological contributions. First, it proposes federated learning frameworks for privacy safe collaborative modeling of structured time series across distributed units, enabling information sharing without direct data exchange. Second, it introduces auxiliary-enhanced learning strategies that exploit contextual information during training to improve prediction robustness when such information is unavailable at deployment, mitigating performance degradation under missing auxiliary inputs. Third, it presents adaptive model architectures capable of processing variable length collections of sequences while maintaining permutation invariance and structural flexibility, allowing seamless adaptation to changing system configurations. These contributions can advance in predictive analytics capabilities for connected systems with applications in Internet of Things (IoT) networks, wearable monitoring, and distributed sensing environments.