On Nonnegative Matrix and Tensor Decompositions for COVID-19 Twitter Dynamics

Authors:
  • LKLara Kassab
    Lara Kassab
  • AKAlona Kryshchenko
    Alona Kryshchenko
  • HLHanbaek Lyu
    Hanbaek Lyu
  • DMDenali Molitor
    Denali Molitor
  • DNDeanna Needell
    Deanna Needell
  • ERElizaveta Rebrova
    Elizaveta Rebrova
Preprint
Publisher website
Google Scholar
Abstract
We analyze Twitter data relating to the COVID-19 pandemic using dynamic topic modeling techniques to learn topics and their prevalence over time. Topics are learned using four methods: nonnegative matrix factorization (NMF), nonnegative CP tensor decomposition (NCPD), online NMF, and online NCPD. All of the methods considered discover major topics that persist for multiple weeks relating to China, social distancing, and U.S. President Trump. The topics about China dominate in early February before giving way to more diverse topics. We observe that NCPD and online NCPD can detect topics that are prevalent over a few days, such as the outbreak in South Korea. The topics detected by NMF and online NMF, however, are prevalent over longer periods of time. Our results are validated against external news sources.