Seminar: PhD seminar: Investigating commuter flows on the London Underground

SpeakerSam Parsons
AffiliationStatistical Science
DateWednesday, 25 Mar 2015
Time13:00 - 14:00
LocationRoom 102, 1-19 Torrington Place
Event seriesStatistical Science Seminars

With over 1.2 billion journeys a year taken on the London Underground, building an accurate model of the passenger distribution over the course of a day is an objective with obvious utility. A Gaussian-Poisson state space model is one option for such a model, but the likelihood function is analytically intractable and most approximation methods are computationally expensive. Using the method of moments to approximately fit the model is shown here to be effective without much computational expense.
Maximum likelihood estimates of model parameters would require the evaluation of the log likelihood function and its gradients, but none of these are analytically tractable. Even a numerical approximation would have parameter dependency whose complexity increased with the length of the data. A composite likelihood approach avoids this problem, so the complexity of the parameter dependence remains constant for all terms in the log composite likelihood. Combining this approach with the method of moments gives an inference framework that is computationally cheap and reasonably accurate.
Furthermore, a low dimensional latent space can be used to place tube stations into clusters. A hierarchical clustering sequence is computed and interpreted in the context of complete station descriptions.
Some of the challenges that come with this methodology, and possible alternatives, are also discussed.

iCalendar csml_id_221.ics