Phase-Change Detection through Dynamic Subspace Learning in Heterogeneous Time-Series Data
PI: Mostafa Reisi Gahrooei, Co-PI: Nicholas Napoli
Award Period: 01/16/2020-10/15/2020
The advances in sensing technology have generated multimodality datasets with complementary information in various domains. For example, in health and human-related applications, a collection of biomarkers gathered over time can be fused to infer phase-changes in the health condition of a patient or cognitive performance of a subject. This collection of co-occurring biomarkers is an example of what we refer to as multimodality datasets. In spite of the wide applications of multimodality datasets, theoretical and algorithmic developments that address how different sources of data can effectively be fused to achieve a more accurate understanding of a system is still in its infancy. Major challenges in the fusion of multimodality datasets include the existence of missing data, the difference in the scale and resolution, and the dynamic interaction of sources of data, to name a few. This project is an effort to address these challenges in integrating multimodality datasets. The focus of this research is on developing novel theoretical and methodological approaches for the fusion of a collection of time-series data with potentially missing values that are co-evolving and interacting over time. More specifically, we expect two major developments. First, we develop a novel self-expressive model that with the aid of roughness and sparsity-induced penalties dynamically clusters the data into cohesive subspaces. Secondly, we will design a system monitoring approach that detects the phase changes in the system. These two steps together assist the understanding of how the system evolves and when the system shifts from one phase to another.