A scalable machine learning approach for measuring violent and peaceful forms of political protest participation with social media data - Fig 8

Above. Time series showing the total presence of social action types in Ferguson, MO the week after the shooting of Michael Brown by Officer Darren Wilson on August 9, 2014. The presence of each action type is determined by our adept Bayes classifier and measured as the sum of posterior probabilities over all tweets from each hour in the plotted span of time. Left. Map of Ferguson, MO depicting clusters of collective force activity over one hour around 12 AM, on August 11th. The size of each cluster-circle represents the area from which tweets emerged (not the number of tweets contained), and the portion of each circle colored red indicates the portion of tweets classified to represent the collective force action. Right. A phrase shift showing the most impactful features present in all tweets classified as being representative of collective force. Phrases on the right pull the classifier toward a positive classification, and phrases on the left pull the classifier towards a negative classification.