Identifying violent protest activity with scalable machine learning ∗

The outbreak and frequency of violent protest activity since 2010 has been a cause for alarm among policy makers and the public at large and has renewed interest in the study of violent forms of protest action. Until recently, the study of violent protest action, and indeed protest action in general has been limited to case studies (Tilly 1988; Tilly and Tarrow 2015), simulation studies (Epstein 2002) and newspaper accounts (Earl et al. 2004). With the widespread use of social media websites such as Twitter and Facebook as means of protest mobilization along with innovations in high dimensional statistics and machine learning researchers are now able to collect large and geographically diverse data for studying protest activity (Barberá et al. 2015; Metzger et al. 2016; Gerbaudo 2012; Tucker et al. 2014). In this paper, we build a series of scalable machine learning algorithms and software which jointly leverage spatial and textual data to identify violent and peaceful protest activity using English language Tweets. We then use our classifier to demonstrate how our software can be used by researchers to construct databases which measure violent and peaceful forms protest activity at fine-grained levels of time and geography and explore relationships between Census demographics and protest activity during the Ferguson protests in November 2015. Finally, we explore how linguistic and spatial features distinguish peaceful from violent forms of collective action. ∗All errors are our own. †ljanastas@uga.edu, http://scholar.harvard.edu/janastas ‡jakerylandwilliams@gmail.com, http://people.ischool.berkeley.edu/%7Ejakeryland/