Moving on Twitter: using episodic hotspot and drift analysis to detect and characterise spatial trajectories

Abstract

Today, a tremendous source of spatio-temporal data is user generated, so-called volunteered geographic information (VGI). Among the many VGI sources, microblogged services, such as Twitter, are extensively used to disseminate information on a near real-time basis. Interest in analysis of microblogged data has been motivated to date by many applications ranging from trend detection, early disaster warning, to urban management and marketing. One important analysis perspective in understanding microblogged data is based on the notion of drift, considering a gradual change of real world phenomena observed across space, time, content, or a combination thereof. The scientific contribution provided by this paper is the presentation of a systematic framework that utilises on the one hand a Kernel Density Estimation (KDE) to detect hotspot clusters of Tweeter activities, which are episodically sequential in nature. These clusters help to derive spatial trajectories. On the other hand we introduce the concept of drift that characterises these trajectories by looking into changes of sentiment and topics to derive meaningful information. We apply our approach to a Twitter dataset comprising 26,000 tweets. We demonstrate how phenomena of interest can be detected by our approach. As an example, we use our approach to detect the locations of Lady Gaga's concert tour in 2013. A set of visualisations allows to analyse the identified trajectories in space, enhanced by optional overlays for sentiment or other parameters of interest.