Probabilistic Models for Dynamic Hypergraphs with Content

Project Overview

The analytical framework for quantifying the uncertainty of outcomes based on large collections of incomplete, and sometimes inaccurate, unstructured information is not currently as mature as its homogeneous counterpart. Bridging this gap is necessary to support uncertainty quantification for applications like graph forecasting and inference with knowledge graphs containing edges based on probable relationships. During this project, we created flexible joint models to relate dynamic hypergraphs and data that are applicable to a variety of datasets and enable real-time discovery and analytics. These probabilistic models allow us to use data to predict content and future social media activity. Experiments on different datasets show that the developed models are also able to learn salient structures within each of these heterogeneous datasets. Additionally, these models successfully address challenging tasks like forecasting, recommendation, and risk analysis.

Mission Impact

Uncertainty quantification is critical when data-driven modeling is applied to high-consequence decision areas. During this project, we developed science and technology tools and capabilities for uncertainty quantification to meet future national security challenges. The developments on this project were successfully transitioned and are now part of a funded effort to extend these developments to knowledge-graph applications. The new models, algorithms, and software developed on this project were shared with the broader research community, other national laboratories, and interested government agencies. We curated multiple datasets that were made available to researchers at the LLNL LLAMA Workshops.

Publications, Presentations, and Patents

Klymko, C. 2021. "Generating Large Graphs Using Vertex Replacement Grammars." Argonne National Laboratory Mathematics and Computer Science Division LANS Seminar. Virtual. July 2021.