PhD Thesis Defense: Iman Deznabi, Adaptive Deep Learning Models for Personalized Modeling of Heterogeneous Time-series Data
Content
Speaker
Abstract
Heterogeneity in time-series data, arising from factors like varying sampling rates, multiple frequencies, multimodal sources, and individual subject uniqueness, poses significant challenges for traditional machine learning models. These models often fail to adapt to new subjects, particularly in limited, sparse, or incomplete data scenarios, highlighting the need for dynamic, personalized models capable of zero-shot and few-shot learning.
This thesis develops adaptive machine-learning models to address the inherent heterogeneity of time-series data. These models dynamically adjust to data characteristics, significantly outperforming standard methods, especially in scenarios with limited data. The research is structured into three parts: (1) introducing dynamic networks for handling heterogeneous data from multiple resolutions and modalities, with applications in ICU mortality prediction, COVID-19 detection, and human activity recognition; (2) developing personalized models for individual variability, demonstrated in stress prediction; and (3) synthesizing these approaches to manage both data-driven and subject-driven heterogeneity, exemplified by an application in zero-shot microclimate prediction.
The proposed models advance predictive modeling and forecasting, with applications across healthcare, environmental science, and beyond. This work highlights the transformative potential of adaptive machine learning in analyzing complex, variable time-series data, paving the way for more accurate and personalized solutions.
Advisor
Ina Fiterau