Please join us for a CSSI seminar, with lunch starting at 12pm, Friday Oct. 10, in our usual room at Lederle Graduate Research Center (LGRC) A112.
Nathan Wycoff, UMass Mathematics and Statistics
Title: Transfer Learning and Emerging Humanitarian Crises
Abstract: The emerging phase of a new humanitarian crisis is often a time when both 1) it is most critical for humanitarian assistance to be delivered, and 2) the least information is available and the dynamics are least understood. Transfer learning, the practice of using plentiful historical data to improve an analysis on a small target dataset, appears to have much to offer in such contexts. However, many existing transfer learning methods assume that historical data are plentiful and/or known to be relevant to the task at hand. This may not be the case in humanitarian crises, such as refugee crises, where historical datasets are incomplete and the relationship between them is unclear. In this talk, we will begin by presenting a case study of predicting the emerging dynamics of the refugee crisis driven by Russia's 2022 invasion of Ukraine using internet data. We will then present methodological work motivated thereby, describing a Bayesian approach to transfer learning which can accommodate small source datasets while performing model averaging over historical dataset relevance.
Bio: Nathan Wycoff is Assistant Professor in Mathematics and Statistics at Umass Amherst. Previously, he was a postdoc at the McCourt School of Public Policy's Massive Data Institute at Georgetown University, where he developed quantitative models of human migration in collaboration with analysts at the United Nations High Commissioner for Refugees. His 2021 PhD in Statistics is from Virginia Tech where he studied sequential experimental design for performing sensitivity analysis of expensive-to-run computer simulations.
