CRII: OAC: Cyberinfrastructure for Machine Learning on Multivariate Time Series Data and Functional Networks

This award is funded in whole or in part under the American Rescue Plan Act of 2021 (Public Law 117-2).

From weather analysis to brain region activity analysis, from traffic flow analysis to financial trend analysis, multivariate time series (MVTS) data have been used extensively in predictive and exploratory tasks for numerous domains. MVTS instances represent states of dynamical systems and natural events using multiple time series of interdependent variables.

Functional networks leverage the interactions of MVTS variables by finding higher-order relationships among them. The appropriate choice of data representation (MVTS or functional network) poses a challenge in machine learning (ML) efforts that can affect the performance of downstream tasks such as classification, regression, and clustering. This project will develop cyberinfrastructure that is public, web-based, and Graphical User Interface (GUI)-enabled and enables both novel and previously developed predictive, exploratory, and generative tasks on both data representations.

The project serves the national interest by promoting the progress of solar physics science through facilitating solar flare prediction from MVTS-based solar magnetic field data and advancing national health through improving prediction models for neurological diseases (e.g., Schizophrenia) from fMRI-based functional brain networks. The research outcomes, including the cyberinfrastructure developed and ML models designed, will provide an opportunity for interdisciplinary research involving undergraduate students including those from underrepresented groups, for course curriculum development, and for high school outreach activities.

MVTS instances are formed from the time series records of multiple sensors. In functional networks, the nodes represent the variables, and the edges represent the statistical similarity of the time series of the corresponding nodes. While in the MVTS representation completeness of data is preserved, noisy or missing data in time series due to events such as faults in sensors can compromise the performance of downstream ML tasks.

Functional network representations help leverage multi-hop relationships of the variables, but the threshold-dependent sparsity in network construction can make ML models lose important features. Machine learning challenges of MVTS and functional network datasets include the appropriate choice of data representation and the limited number of training samples (especially in the medical domain).

This project will provide a unified framework for performing (un)-supervised ML tasks on both data representations through application and customization of contemporary ML modes such as matrix/tensor decomposition, sequence models, Graph Neural Networks (GNN), and dynamic graph embedding. The project will also provide a framework for augmenting datasets with synthetic training samples through autoregressive, autoencoder-based, and adversarial models.

The project will design and implement a web-based system that contains modules for data import and preparation, representation learning, data augmentation, validation, result visualization, and for exporting derived and synthetic datasets. The web platform will be hosted in the public domain, and its GUI-based front end will enable researchers to apply back-end ML models without explicitly programming using ML libraries.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.