Data Science and Statistics Seminar by Dr. Patrick Brandt (UT Dallas EPPS)
From publicly posted calendar Checked yesterday
About this event
Title: Forecasting Accurately in Space and Time (FAST): Explainability for End-Users Abstract: Quantitative forecasts of conflict and violence should be understandable for end-users, including governments, IOs, NGOs, journalists, and private enterprise. In this paper, we report our efforts at explanation in the context of a policy transfer activity: the Fast Forward: Forecasting Global Emerging Threats Competition. Along with conflict forecasts, this competition required teams to submit machine-generated explanations for each forecast, to be assessed by intelligence analysts. An explanation interface must do several things to effectively inform analysts, but most importantly we argue that explanations should contextualize the forecast, recognize its limits, and understand its applicability for a purpose. We found value in four elements of explanations: (1) textual bottom-line-up-front summaries; (2) graphical representations of time trends alongside forecasts; (3) graphical representations of input data, including where the geographic unit sits in the global distribution; and (4) comparisons of the forecast to the global set of forecasts. We explain the development of our underlying models and the format and processes for generating explanations and visualizing results. This serves as an initial template for others to replicate, build on, and modify for their own purposes. Authors: Scott Robert Patterson, Patrick T. Brandt, Vito D’Orazio, Max Z. Margulies and Theodore McLauchlin