Geospatial Science Seminar 13.03.2012

A kernel based approach for spatio-temporal modelling and forecasting.
James Haworth, UCL Department of Civil, Environmental and Geomatic Engineering.

To download a PDF of the seminar please click here.

Abstract.
Traditionally, statistical models have been used for spatio-temporal forecasting due to their strong theoretical foundation and interpretability. However, many large scale spatio-temporal datasets display complex, nonlinear, nonstationary properties that violate the iid assumptions of classical statistical models. Increasingly, practitioners are borrowing techniques from the machine learning community because of their innate ability to deal with this type of data. In particular, kernel based approaches such as the support vector machine have been successful because they use the so called “kernel trick” to allow linear algorithms to model nonlinear data. In this session, a kernel based approach to spatio-temporal forecasting is introduced. The model is tested using travel time data collected by automatic number plate recognition cameras on London’s road network.