Date: Friday, December 12, 2014
Location: Rockefeller Hall 310
Title: A taste of data mining: predicting ambulance demand
In this talk, I will introduce an interesting data mining project. The goal is to predict ambulance demand for the city of Toronto, Canada. To be useful for decisions regarding staff and fleet management and dynamic deployment, we need to predict this demand accurately for every two hour time period and every 5x5 city block region. There are several challenges: although the dataset is huge, the number of observations per time period and locality is almost always 0. There are also complex patterns in the demand across time and space that we need to take into account and take advantage of. I will propose a new statistical method to address these challenges. I will motivate why this new method produces 20-25% more accurate predictions than the current EMS industry practice.
Throughout the talk I will point out connections to common themes in data mining (typical modeling process, typical challenges and their solutions, how to evaluate outcomes, etc.), and to aspects of some other sub-fields of applied mathematics (operations research, simulation, time series, Bayesian statistics, optimization, etc.). Most of the talk is accessible to a general mathematical or scientific audience and will assume no prior knowledge of statistics.