Measuring coordinated vs. spontaneous activity in online social movements

Abstract

Social media platforms provide people all over the world with an unprecedented ability to organize around social and political causes. However, these same platforms enable institutional and organized actors to engineer fabricated social movements to advance their agenda. These “astroturfing” or “false amplification” phenomenons leverage a variety of of tools and techniques, ranging from fully automated bot activity to accounts manned by extrinsically motivated (e.g. compensated) human operators. These campaigns also range from simple spam operations to sophisticated efforts involving numerous orchestrated accounts, sometimes coordinated across linguistic and cultural clusters. While the former category is straightforward to analyze via data mining methods, sophisticated fabricated campaigns in the latter category are engineered to mask their true nature from the public.Working from the proposition that a large number of accounts controlled by a small number of coordinated entities will lack the behavioral diversity of a similar number of accounts controlled by uncoordinated individual actors, we propose a framework of signals (metrics) along three dimensions: Network: how accounts are connected to one another, and the clusters they form within the online conversation, Temporal: patterns of messaging across time in the online conversation, Semantic: an observation of the diversity of topics and meaning throughout the online conversation.We test this framework on three case studies: the online conversation on #ColumbianChemicals in the U.S., the international discussion of the #DopingLeaks event, and an analysis of political discussions in Venezuela. In all three cases, we find what we assess to be anomalous, fabricated behavior on at least one dimension.

Trends in Detection and Characterization of Propaganda Bots

“…To do so would require either a large number of bots acting in an abnormally coherent fashion, or a smaller number of bots tweeting at abnormally high rates. As we have seen, [1], [23], [24] both of these tactics produce temporal signatures that are detectable to some degree by state of the art classifiers.…”

“…Because propaganda botnets are attempting to draw attention to their topic, they frequently act in temporally correlated ways that create a detectable signature. [1], [14], [23], [24]. Other behavioral features include following circadian cycles for appropriate time zones, diversity in rate and topic of tweets, etc.…”

“…Further, the keywords and hashtags must be updated for each emerging topic of study. Still, it is common practice to filter based on topical keywords when assembling an initial data set associated with a particular political event or movement [23].…”

“…Computational propaganda is no different. There are at least three distinct generations of social bots in existence at this time [23]:…”

“…However, there is a limit to how far an influence bot can disguise itself and still have the desired impact [23]. In order to promote a topic vociferously enough to ensure that the ranking algorithms elevate that to a trending topic, a propaganda botnet must create a sufficiently high number of messages in a short amount of time.…”