Onur Varol

The rise of social bots

Todays social bots are sophisticated and sometimes menacing. Indeed, their presence can endanger online ecosystems as well as our society.

BotOrNot: A System to Evaluate Social Bots

While most online social media accounts are controlled by humans, these platforms also host automated agents called social bots or sybil accounts. Recent literature reported on cases of social bots imitating humans to manipulate discussions, alter the popularity of users, pollute content and spread misinformation, and even perform terrorist propaganda and recruitment actions. Here we present BotOrNot, a publicly-available service that leverages more than one thousand features to evaluate the extent to which a Twitter account exhibits similarity to the known characteristics of social bots. Since its release in May 2014, BotOrNot has served over one million requests via our website and APIs.

Evolution of online user behavior during a social upheaval

Social media represent powerful tools of mass communication and information diffusion. They played a pivotal role during recent social uprisings and political mobilizations across the world. Here we present a study of the Gezi Park movement in Turkey through the lens of Twitter. We analyze over 2.3 million tweets produced during the 25 days of protest occurred between May and June 2013. We first characterize the spatio-temporal nature of the conversation about the Gezi Park demonstrations, showing that similarity in trends of discussion mirrors geographic cues. We then describe the characteristics of the users involved in this conversation and what roles they played. We study how roles and individual influence evolved during the period of the upheaval. This analysis reveals that the conversation becomes more democratic as events unfold, with a redistribution of influence over time in the user population. We conclude by observing how the online and offline worlds are tightly intertwined, showing that exogenous events, such as political speeches or police actions, affect social media conversations and trigger changes in individual behavior.

Traveling trends: social butterflies or frequent fliers?

Trending topics are the online conversations that grab collective attention on social media. They are continually changing and often reflect exogenous events that happen in the real world. Trends are localized in space and time as they are driven by activity in specific geographic areas that act as sources of traffic and information flow. Taken independently, trends and geography have been discussed in recent literature on online social media; although, so far, little has been done to characterize the relation between trends and geography. Here we investigate more than eleven thousand topics that trended on Twitter in 63 main US locations during a period of 50 days in 2013. This data allows us to study the origins and pathways of trends, how they compete for popularity at the local level to emerge as winners at the country level, and what dynamics underlie their production and consumption in different geographic areas. We identify two main classes of trending topics: those that surface locally, coinciding with three different geographic clusters (East coast, Midwest and Southwest); and those that emerge globally from several metropolitan areas, coinciding with the major air traffic hubs of the country. These hubs act as trendsetters, generating topics that eventually trend at the country level, and driving the conversation across the country. This poses an intriguing conjecture, drawing a parallel between the spread of information and diseases: Do trends travel faster by airplane than over the Internet?

Clustering memes in social media

The increasing pervasiveness of social media creates new opportunities to study human social behavior, while challenging our capability to analyze their massive data streams. One of the emerging tasks is to distinguish between different kinds of activities, for example engineered misinformation campaigns versus spontaneous communication. Such detection problems require a formal definition of meme, or unit of information that can spread from person to person through the social network. Once a meme is identified, supervised learning methods can be applied to classify different types of communication. The appropriate granularity of a meme, however, is hardly captured from existing entities such as tags and keywords. Here we present a framework for the novel task of detecting memes by clustering messages from large streams of social data. We evaluate various similarity measures that leverage content, metadata, network features, and their combinations. We also explore the idea of pre-clustering on the basis of existing entities. A systematic evaluation is carried out using a manually curated dataset as ground truth. Our analysis shows that pre-clustering and a combination of heterogeneous features yield the best trade-off between number of clusters and their quality, demonstrating that a simple combination based on pairwise maximization of similarity is as effective as a non-trivial optimization of parameters. Our approach is fully automatic, unsupervised, and scalable for real-time detection of memes in streaming data.

Predicting Online Extremism, Content Adopters, and Interaction Reciprocity

We present a machine learning framework that leverages a mixture of metadata, network, and temporal features to detect extremist users, and predict content adopters and interaction reciprocity in social media. We exploit a unique dataset containing millions of tweets generated by more than 25 thousand users who have been manually identified, reported, and suspended by Twitter due to their involvement with extremist campaigns. We also leverage millions of tweets generated by a random sample of 25 thousand regular users who were exposed to, or consumed, extremist content. We carry out three forecasting tasks, (i) to detect extremist users, (ii) to estimate whether regular users will adopt extremist content, and finally (iii) to predict whether users will reciprocate contacts initiated by extremists. All forecasting tasks are set up in two scenarios: a post hoc (time independent) prediction task on aggregated data, and a simulated real-time prediction task. The performance of our framework is extremely promising, yielding in the different forecasting scenarios up to 93% AUC for extremist user detection, up to 80% AUC for content adoption prediction, and finally up to 72% AUC for interaction reciprocity forecasting. We conclude by providing a thorough feature analysis that helps determine which are the emerging signals that provide predictive power in different scenarios.

Clustering memes in social media streams

AbstractThe problem of clustering content in social media has pervasive applications, including the identification of discussion topics, event detection, and content recommendation. Here, we describe a streaming framework for online detection and clustering of memes in social media, specifically Twitter. A pre-clustering procedure, namely protomeme detection, first isolates atomic tokens of information carried by the tweets. Protomemes are thereafter aggregated, based on multiple similarity measures, to obtain memes as cohesive groups of tweets reflecting actual concepts or topics of discussion. The clustering algorithm takes into account various dimensions of the data and metadata, including natural language, the social network, and the patterns of information diffusion. As a result, our system can build clusters of semantically, structurally, and topically related tweets. The clustering process is based on a variant of Online K-means that incorporates a memory mechanism, used to “forget” old memes and replace them over time with the new ones. The evaluation of our framework is carried out using a dataset of Twitter trending topics. Over a 1-week period, we systematically determined whether our algorithm was able to recover the trending hashtags. We show that the proposed method outperforms baseline algorithms that only use content features, as well as a state-of-the-art event detection method that assumes full knowledge of the underlying follower network. We finally show that our online learning framework is flexible, due to its independence of the adopted clustering algorithm, and best suited to work in a streaming scenario.

Distilling the Outcomes of Personal Experiences: A Propensity-scored Analysis of Social Media

Millions of people regularly report the details of their real-world experiences on social media. This provides an opportunity to observe the outcomes of common and critical situations. Identifying and quantifying these outcomes may provide better decision-support and goal-achievement for individuals, and help policy-makers and scientists better understand important societal phenomena. We address several open questions about using social media data for open-domain outcome identification: Are the words people are more likely to use after some experience relevant to this experience? How well do these words cover the breadth of outcomes likely to occur for an experience? What kinds of outcomes are discovered? Studying 3-months of Twitter data capturing people who experienced 39 distinct situations across a variety of domains, we find that these outcomes are generally found to be relevant (55-100% on average) and that causally related concepts are more likely to be discovered than conceptual or semantically related concepts.

What is gained and what is left to be done when content analysis is added to network analysis in the study of a social movement: Twitter use during Gezi Park

ABSTRACT As social movements relying on the weak ties found in social networks have spread around the world, researchers have taken several approaches to understanding how networks function in such instances as the Arab Spring. While social scientists have primarily relied on survey or content analysis methodology, network scientists have used social network analysis. This research combines content analysis with the automated techniques of network analysis to determine the roles played by those using Twitter to communicate during the Turkish Gezi Park uprising. Based on a network analysis of nearly 2.4 million tweets and a content analysis of a subset of 5126 of those tweets, we found that information sharing was by far the most common use of the tweets and retweets, while tweets that indicated leadership of the movement constituted a small percentage of the overall number of tweets. Using automated techniques, we experimented with coded variables from content analysis to compute the most discriminative tokens and to predict values for each variable using only textual information. We achieved 0.61 precision on identifying types of shared information. Our results on detecting the position of user in the protest and purpose of the tweets achieved 0.42 and 0.33 precision, respectively, illustrating the necessity of user cooperation and the shortcomings of automated techniques. Based on annotated values of user tweets, we computed similarities between users considering their information production and consumption. User similarities are used to compute clusters of individuals with similar behaviors, and we interpreted average activities for those groups.

Spatiotemporal analysis of censored content on Twitter

Social media have become vehicles for instantly disseminating and accessing information on a global scale. Beside such positive contributions, social media also enable malicious activities such as recruitment for terrorist groups or coordinate orchestrated campaigns. Censorship is one way of limiting user activities, but applying it fairly is not easy, as exemplified by site-blocking censorship by governments. To avoid complete site-blocking, some social media sites have complied with requests by governments for content removal or partial censorship. In this study, we analyzed our collection of more than 100,000 tweets that were either censored tweets or retweeted censored content. We showed variability in audience location using time zone and language preferences as proxies, which is not bounded by geographic location of the censorship. We show that most of the time content find its way to disseminate and reach broader audience even with the censorship.