In this month’s research spotlight, COSMOS highlights recent research that looks at a multi-theoretic framework to simulate mobs. Specifically, the study authored by Dr. Samer Al-Khateeb and Dr. Nitin Agarwal titled, “Mob Simulation Guided by Social Science-based Multi-Theoretical Framework,” uses the Monte Carlo method to simulate the formation of mobs and evaluate their success. The study was published at the 33rd European Conference on Operational Research (EURO 2024) in the stream on ‘Behavioural Operation Research (OR)’ in the session on ‘Modelling social-behavioural phenomena in creative societies’ during June 30 – July 3, 2024, held at Copenhagen, Denmark.

“Social media sometimes empowers potential adversarial state actors, paid trolls, and

extremists to incite hysteria and coordinate nefarious actions, e.g., deviant mobs,” they explain. “A mob is an event organized via social media or other forms of digital communication technologies in which a group of people gathers online and/or offline to conduct an act collectively and then disperses.” In their presentation of the research, they outlined the mob organization process, which comprises five phases: planning, recruitment, execution, replaying/republishing, and evaluation.

After reviewing the literature, they categorized mobs into four main categories: based upon environment, actors, time, and motives. For example, mobs can be differentiated by environment, such as cyber versus physical mobs. Similarly, entertainment mobs are differentiated from deviant or malicious mobs, as are flash mobs from long mobs. With the advent of AI and bot accounts, mobs can also be differentiated into human actors versus automated actors, or involve a hybrid of the two. Numerous real-world examples show these different categories. Some they mentioned were entertainment flash mobs that occurred in Seattle, Japan, and Thailand; an animal rights protest mob against Burger King breeding practices in Paris; a “flash investing” mob, where Reddit-coordinated investments skyrocketed GameStop stock value; and the Washington DC State Capital riot mob, among several others. 

These categorizations matter not only to interpreting the effect of such mobs, but also in the  prediction of formation. For example, social bot mobs can be predicted based on differences in metadata between human and automated or bot-based networks. Social bots tend to share location and language less than humans, and bot communication networks generally are less divided (i.e., having fewer components) and less densely connected between those components than in human-based communication networks.

Dr. Agarwal and Dr. Al-Khateeb also looked to past NATO exercises that looked at how to track cyber mobs. Namely, they analyze the components of the NATO 2015 Trident Juncture Exercise, where Social Cyber Forensics (SCF) were used to track connections between different blogs and the users who accessed them. By discovering the metadata of users that accessed blogs—data such as IP addresses, names associated with registered domain names, email addresses, phone numbers, or other factors of digital presence like Facebook, Twitter, or YouTube accounts—relationships between blogs could be established or new, unknown influential blogs discovered.

Having thus established the factors key in considering successful real-world mobs, Dr. Agarwal and Dr. Al-Khateeb then went on to describe the structure of their mob simulation. Their simulations were based upon a framework known as the Monte Carlo method, which considers three factors that affect the decision making of individual mobbers: interest, control, and power. Under this framework, four scenarios are considered based upon interest and control: 1) if a mobber has interest and control, then the mobber will likely act; 2) if a mobber has interest but no control, then the mobber has a 50/50 percent chance to act or withdraw; 3) if a mobber has control but no interest, then the mobber has a 50/50 percent chance to withdraw or “power exchange” (a term which denotes relinquishing power to another in exchange for social power or recognition); or finally 4) if a mobber has no interest nor control, then they have a 50/50 percent chance to withdraw or act against the mob group itself.

These four scenarios then factor into an equation calculating the participation rate, which is a percentage calculated based on the number of mob participants out of the total number of invited individuals. This equation has two forms: one in which the number of actors against the mob are considered, and one where no acts against the mob are considered. Essentially, there is a slightly different equation when considering the mob as a single event (where there is no counteraction) versus considering it as two competing sides or events (i.e. “pro” versus “anti” actors). Once calculated, the participation rate is considered for evaluating whether a mob was successful or not, with a threshold value needing to be achieved for success.

There were three research questions that they sought to answer using their simulations, given the chosen parameters (i.e., a certain number of invited people, a certain number of powerful actors, a certain number of simulations, and a certain threshold value): 

  1. What is the chance a given mob will succeed?
  2. How many powerful actors are needed to have a successful mob?
  3. How does the time it takes for an invited participant to decide to join a mob affect the mob participation rate?

They ran the simulation multiple times with different numbers of individuals, powerful actors, and thresholds of success. Each time they ran a simulation, each invited individual had interest and control randomly assigned, and the simulation then calculated whether the mob resulted or failed and then aggregated results (such as the average participation rate and total number of successes for that given set of parameters). They grouped simulations into three primary experiments: one with zero powerful actors and a set number of invited people (to discover what threshold values would be necessary for a mob with zero or no known powerful actors to succeed); one with variant numbers of both invited people and powerful actors (to discover how the number of people involved and the number of powerful actors were correlated); and one where control and interest were randomly determined in timed waves (to discover how a mob might be affected by the decision-making timing of mobbers).

Dr. Agarwal and Dr. Al-Khateeb came away with three primary conclusions from the simulations. First, individuals who acted against the mob had a more negative effect on the mob’s success than those who withdrew or exchanged power. Second, larger mobs needed more powerful actors to be successful. Finally, the participation rate increased when mobbers were simulated in cascades or waves.

They concluded their research presentation by summarizing the next steps necessary in continuing this research: improving the theoretical and cognitive models, validating the models with robust ground truth data, and then validating the new simulations with real-world data. 

Dr. Agarwal said,  “This critical research advances our understanding of mobs in digital and physical spaces. Our methodology provides a theoretically grounded and systematic approach to simulate mobs and evaluate conditions conducive to a successful mob. The research aims to provide capabilities in the hands of decision makers to proactively manage mobs.”