Applied Math Seminar

Fall 2024

All talks are from 12:00-1:00 p.m. in the Seminar Room CH351, unless otherwise specified.

• Dec

  03
• TBA

  Ron Malone

  USNA-Physics

  Time: 12:00 PM

• Nov

  19
• Active preference learning from paired comparisons in ordinal embeddings

  Greg Canal

  Johns Hopkins-APL

  Time: 12:00 PM

  View Abstract

  Paired comparisons are a classic and effective modality to query the preferences of human responders by having them provide ranking responses over pairs of objects, such as "I prefer object A over object B." Learning predictive preference models from paired comparisons underlies many modern machine learning frameworks including recommender systems and finetuning language models. However, in general is in infeasible to query individual humans with a large number of comparisons to learn their preferences, necessitating the development of "active learning" methods to intelligently select an informative subset of query comparisons. In this talk, I will discuss computationally efficient, information-theoretic strategies to intelligently sample paired comparisons in ordinal embedding spaces, as well as learning methods that simultaneously learn user preferences along with the appropriate embedding space.

• Nov

  12
• Long-Term Viewership Forecasting of American College Football Games

  Junhyung Park

  USNA-Math

  Time: 12:00 PM

  View Abstract

  Using data collected from televised college football games between 2014–2019, we present a log-linear statistical model for viewership of NCAA college football games that controls for well known factors such as temporal fixed-effects, strength of the games, rivalries, outcome uncertainty and broadcast medium, among others. Novel factors influencing viewership that are not used in previous studies include the number of concurrent broadcasts, the strength of the game relative to other concurrently broadcast games and their interactions. This model only includes team-specific factors available prior to the season thereby providing valuable input if games were to be scheduled well in advance. We also propose a novel variable coding that allows a parsimonious estimation of the effect of all pairs of inter-conference games, not just intra-conference games. The model is assessed through residual analysis and in- and out-of-sample predictive performance.