Exploring Population Dynamics with Crowd-Sourced Data

Increasingly large volumes of geo-located data from social messaging are available in the public domain. These data contain valuable clues about daily activity patterns. For some of the more prolific users it is possible to trace sequences of activity with considerable spatial and temporal detail.

In this paper, we use a substantial sample of twitter data from the city of Leeds to investigate space-time activity patterns. A framework of activity types will be presented, and we will evaluate different methods for the classification of behaviour based on message content. These methods range from manual calibration by an intelligent observer through text recognition to machine learning. Relevant metrics will be proposed and tested for the automated procedures which will be necessary in order to process data of this type in useful volumes. We will also present substantive results which illustrate variations in activity patterns across the city at different times in the day, and consider the potential for examination of individual sequences and movement patterns. We will offer some thoughts on both the ethical robustness of this approach, and its potential for generalisation beyond a self-selecting sample of twitter users. The potential for model-based approaches built on the foundations of this analysis will be considered.