In the study of time-series data, the presence of heterogeneity presents a complex challenge for traditional machine learning and deep learning models. This variability can stem from a multitude of sources, such as differences in sampling rates, the presence of multiple frequencies of interest, and the multimodal nature of data collected from various origins. Additionally, the uniqueness of individual subjects further compounds the complexity, introducing variability that requires special personalized models to handle. Given these challenges, conventional models often fall short when tasked with analyzing this type of data, because they fail to personalize the architectures to individuals, particularly when the data is limited, sparse or incomplete highlighting the need for dynamic models that can personalize to new subjects using minimal data in a few-shot or zero-shot manner.
The proposed thesis focuses on the development and application of adaptive machine-learning models designed to navigate and leverage the heterogeneity inherent in time-series data. These models are built with the flexibility to adjust their parameters and structures in response to the specific characteristics of the data they encounter, offering a tailored approach that significantly outperforms standard and state-of-the-art models, particularly in limited data scenarios.
The work is structured in three main parts, each targeting a distinct aspect of the problem. Initially, we explore methods to address the intrinsic data characteristics, introducing dynamic networks capable of integrating data from multiple sources, each with its unique resolution and modality. This section demonstrates the enhanced capability of these models to process and analyze heterogeneous data effectively, showing notable improvements in benchmarks belonging to different applications such as ICU in hospital mortality prediction, early COVID-19 prediction, human activity detection and wearable affective detection.
Subsequently, the focus shifts to personalizing models to accommodate the differences among individual subjects. This part of the research presents innovative approaches for personalized prediction, showcasing the models' effectiveness in scenarios ranging from zero-shot learning to online prediction and forecasting. The applications explored include stress prediction and climate forecasting, areas where personalization and adaptability are crucial for accurate model performance.
To close the methodological gap , we plan to develop dynamic methods that synthesize the approaches from the first two parts, proposing models that adeptly manage both the data-driven and subject-driven heterogeneities. These comprehensive models represent a significant advancement in the field, capable of dynamically and efficiently addressing the dual challenges presented by heterogeneous time-series data.
By bridging the gap between the diverse characteristics of time-series data and the need for personalized analysis, this research contributes valuable insights and methodologies to the domains of predictive modeling and forecasting. The adaptability and versatility of the developed models not only enhance their utility across a broad spectrum of fields, including healthcare and environmental science but also pave the way for future innovations in the analysis of complex, variable data streams. Through this work, we underscore the potential of adaptive machine learning to transform our understanding and utilization of heterogeneous time-series data, providing a foundation for more accurate, robust, and personalized data analysis solutions.
Advisor: Ina Fiterau