Ongoing projects in the lab, spanning multilinear subspace learning, graph‑based forecasting, and covert network analysis.
Building a National Security Center of Excellence around three research pillars — advanced manufacturing and munitions (Build), resilient operational technology and industrial control systems (Protect), and secure defense supply chains (Sustain) — woven together by a shared AI/ML digital core. Part of the South Dakota Center for Advanced Research in Materials, Operations, and Resilience (SD‑ARMOR).
Advancing data‑driven identification of nonlinear dynamics from high‑dimensional spatiotemporal data — extending Sparse Identification of Nonlinear Dynamics (SINDy) to tensor‑structured data, establishing rigorous identifiability results, and linking tensor‑based system identification to Koopman‑invariant subspaces.
Extending our KOMET and SLATE temporal domain generalization frameworks (ICMLA 2026) to dynamic systems and control: rather than continuously re‑solving for a feedback gain as a plant drifts over time, we forecast the gain trajectory itself via Koopman/EDMD identification, enabling predictive gain scheduling with no online redesign.
As part of a NASA‑funded, multi‑institution effort fusing SAR and multispectral satellite time series for deep‑learning soil moisture retrieval, our contribution focuses on tensor factorization methods and Gaussian‑process priors for regularizing and interpreting the retrieval models.
This project aims to extend classical ARMA‑like time‑series models to a multilinear framework for forecasting n‑dimensional structured data — images, video, and dynamic graphs.
Investigating how high‑order tensor structures can extend shallow and deep learning algorithms for pattern analysis and knowledge discovery.
Video surveillance systems enable investigators to identify the time, place, and context of significant events, but current methods to sort through the video are labor‑intensive and time‑consuming. Automating the extraction of significant events lets investigators spend more time on other critical tasks, and provides a mechanism to notify investigators while significant events are occurring so they can deploy rapid‑response teams before it's too late to resolve the incident. Video from security cameras is often low‑quality because of poor lighting or obstructions in the scene; we explore a variety of techniques to automatically identify significant motion and enable investigators to efficiently interact with the data. We found that multiple object‑tracking solutions can rapidly identify motion with few false negatives — techniques that can be combined into a distributed system to efficiently present investigators with the information they need in a digestible form.
Social networks are rarely flat — members belong to nested communities such as partners, organizations, and ideologies at once. This project extends graph convolutional networks to capture that full hierarchy rather than just a node's immediate neighborhood, by allowing convolutions to skip across hierarchy levels and assembling the resulting multi‑scale embeddings into a tensor, improving node classification and link prediction on real‑world social networks.
Extends low-rank matrix factorization for time-series forecasting to multilinear (tensor-valued) data — video, dynamic networks, and geospatial grids — by jointly learning a low-rank spatiotemporal factorization and forecasting within the compressed latent space via a transform-based tensor autoregression. The framework handles missing and nonstationary observations, outperforming matrix-based baselines across energy, climate, video, and traffic datasets.
Integrating disparate multimodal data to build accurate models of covert networks commonly found in illicit markets — trafficking, counterfeit goods, critical minerals, and more.
Randy C. Hoover