Gaussian Processes for Geo-Temporal series.
Exploration of the application of Gaussian Processes to Geo-Temporal series with focus on crime distribution forecast in a city.
Requirements
- M.Sc. in Machine Learning, Computer Science, Mathematics, Physics, or similar
- Basic knowledge of time sries
- Decent familiarity with statistical and Baysian methods
- Familiarity with Python and software develoment
Description
Predicting crimes is crucial for law enforcement agencies (LEAs) to optimally allocate resources with the scope to better respond to criminal activities. In this sense, it is essential to forecast the possible crime hotspots within narrow regions spatially by developing a geo-temporal crime forecasting model that can capture crime incidents’ spatial and temporal dependencies.
Gaussian Processes (GPs) are particularly suited to adress this problem due to their ability to model complex spatial and temporal dependencies, inherent in crime data. GPs, not only offer a non-parametric approach that can adapt to various data distributions and incorporate uncertainty estimation, but also allow to explicitly model spatio-temporal correlations directly into the covariance function (kernel), making them interesting candidates to model geo-temporal processes. This thesis comprises three main steps: (1) gain familiarity with Gaussian Processes for time series; (2) litterature review of GPs applied to geo-temporal problesm, specifically to crime forecasting; (3) designing novel GP tailored to the use case of Boston crime dataset.