The Influence Team seeks to understand and characterize influence in a variety of environments such as campaigns using social media, campaigns focused on a country’s leadership, radicalization efforts, and local sources of influence.Team Leads: Paul Narula, Christine Brugh, Rob Johnston
Social Sifter is a project focused on detection of foreign state-sponsored social media influence campaigns. Data from a variety of platforms like Twitter, Reddit, and YouTube are brought into a standard format, pre-processed and augmented with built in human and machine-learning derived heuristics. After data ingest, the system then executes a series of machine learning models and algorithms which test for the existence of a foreign disinformation influence campaign and the presence of authentic and inauthentic users contributing to those campaigns. At present, models undergoing testing include Motif models (which are time series models which focus on network attributes) as well as text models, metadata models, Functional Data Analysis, and other time-series analyses. The Social Sifter system now also has over thirty manually curated test cases and dozens of campaigns that have undergone testing. Additional videos of this work cover our models in greater depth.
Participants: Rob Johnston (Johnston Analytics), Clint Watts (Miburo Solutions), Brent Younce (LAS)
Inorganic users or Bots are often used for spreading misinformation in social media as a part of influence campaigns. As a result, quick identification of Bots is an important task. This project constructs some key feature vectors through statistical analysis and develops algorithms for detection of Bots in both static and dynamic environments with very high accuracy. Results from a preliminary analysis on characteristics of within-Bot networks are also reported - this will potentially provide new insights into interactions among different sub-groups of Bots under types of influence campaigns.
Participants: Dhrubajyoti Ghosh (NCSU), William Boettcher (NCSU), Rob Johnston (Johnston Analytics), Soumendra Lahiri (WUSTL)
Social media provides a far-reaching platform for campaigns of all types, including political and marketing campaigns, to quickly disseminate information and beliefs to a widespread audience. Savvy political influencers have been able to utilize these online platforms to successfully promote misinformation campaigns. However not all campaigns are malicious nor successful in gaining traction and reaching their audience. We present new methods to classify and predict social media campaigns. The social media campaigns are modeled through a functional data analysis via a Multilevel Functional Principal Component Analysis. This model separates the campaign-specific variation from the mean trend, while accounting for additional covariate effects. In addition to accurately estimating the daily information diffusion of the topic, these methods result in coefficients that are used as covariates in machine learning methods to classify campaigns and predict future behavior. These methods we applied to previously labeled data and were able to accurately classify the campaigns. The prediction analyses were able to determine characteristics to determine the future success of the campaigns. These methods are implemented onto the Social Sifter Platform.
Participants: William Rand (NCSU), Anthony Weishampel (NCSU)
To examine posting patterns associated with influence in online health misinformation, our project focused on online discussions of the use of hydroxychloroquine to treat patients with Covid-19. We identified influential pro- and anti-hydroxychloroquine users via a dataset from Mutlu et al., which contained tweets discussing the efficacy of hydroxychloroquine for treating Covid-19. Each tweet had been hand-labeled with the stance it took on the drug. After collecting up to 3200 of each of the identified users’ most recent tweets, we examined both the temporal and topical patterns associated with influence within the different stances. Influential pro-hydroxychloroquine users were found to post more often on average than both influential anti-hydroxychloroquine users and less influential pro-hydroxychloroquine users. We analyzed the contents of the tweets using the GSDMM text clustering method (Yin & Wang, 2016). The GSDMM method identified 58 topics within our corpus of tweets. Similar to the temporal analysis, we discovered that influential pro-hydroxychloroquine users displayed differing posting patterns when compared to influential anti-hydroxychloroquine and non-influential pro-hydroxychloroquine users. In particular, these influential pro-hydroxychloroquine users had fewer tweets associated with interpersonal interactions than users in the two other groups. To incorporate these temporal and topical patterns into a single model, we have fit a mixed Markov model (MMM) to the tweets of each of the three groups. In our ongoing work, we are applying this MMM framework to create a classifier to predict which users are likely to become influential.
Participants: William Rand (NCSU), Iris Bennett (NCSU)
Understanding state-sponsored influence on social media has been a prevalent problem since the 2016 election cycle. Network analysis, specifically through motifs, allows for the classification of networks of accounts as belonging to a state-sponsored influence campaign. This method is applied to Twitter data, with an example application presented. We conclude with a discussion of future applications.
Participants: Khuzaima Hameed (NCSU)
As the U.S. returns to its traditional focus on near-peer competitors and “great power” struggles for influence and allies, academics and policymakers are reexamining elite-focused influence campaigns (covert and overt) to cultivate regional elites. While once a standard element of Cold War projects in Southeast Asia, Africa, and Latin America; these efforts were often unsuccessful, subject to counter-campaigns by regional adversaries, and subordinate to “hard power” measures like military aid and training programs. As the U.S. turned to counter-insurgency and counter-terrorism operations after 9/11, both China and Russia continued to engage in sophisticated campaigns to develop “friendly” local governments. China’s Belt and Road Initiative (BRI) focused its economic wealth and government dominance of industry on massive development projects, tying the fate of local elites to Chinese largesse and allowing China access to markets, resources, and occasionally military bases around the world. Not to be outdone, Russia has similarly made inroads in Central Africa and Latin America, working with governments with dubious human rights records and providing access to technologies often restricted by the U.S. This project reviews Cold War “elite capture” campaigns (both covert and overt) by the “great powers” and compares them to current campaigns by China and Russia. It pays particular attention to new elements of these campaigns that are attuned to the modern economic, political, and social contexts in which regional elites are embedded. This project includes case studies of influence campaigns in Central African Republic, Nigeria, Panama, and Venezuela. These case studies are supplemented by practitioners interviews conducted by the research team.
Participants: Bill Boettcher (NCSU), Rob Johnston (Johnston Analytics), Soumendra Lahiri (WUSTL), Margaret Harney (NCSU), Natalie Kraft (LAS), Dhrubajyoti Ghosh (WUSTL), Caitlyn Roykovich (NCSU), Kate Sanborn (LAS), Tuhin Majumder (NCSU)
This video briefly summarizes our work on Terrorist Membership and Radicalization with the Computational Social Science group over the last year. We briefly introduce three different projects that we have completed: (1) Update of a Systematic Review, (2) Creation of a new dataset through a Data Merge, and (3) creation of a sociopolitical event timeline. For each of these project we provide a brief context of the importance of this work and identify our anticipated next steps for each of those projects.
Participants: Sarah Desmarias (NCSU), Joseph Simons-Rudolph (NCSU), Christine Brugh (NCSU), Ian Siderits (NCSU), Alexa Katon (NCSU), Peyton Frey (NCSU), Kate Sanborn (NCSU)
This research describes the underlying theory behind synthetic population methodology. Using a dataset of known terrorism-involved individuals, this work applies the radicalization use case of estimating the accuracy and precision of identifying informants, criminals, and terrorists.
Participants: Felecia Morgan-Lopez (LAS)
Imagine an interactive capability that immediately alerts you to real-time threat events worldwide. Imagine a capability that enables you to take advantage of the treasure trove of publicly available information to glean insight into the internal posture of activities you deem of interest. This posture would present information from the local perspective; by selectively amassing images, video, audio and text; then organize and triage that data; apply analytics to extract the information of interest; and finally, alert the user to additional related activities. On-the Ground (OTG) is a prototype project delivering such capabilities!
Participants: Felecia Morgan-Lopez (LAS), John Slankas (LAS), Tina Kohler (LAS), Zanetta Tyler (NC A&T), Sheila Bent (LAS), Jenaye Minter (LAS)