Modeling the Influence of Narratives on Collective Behavior Case Study: Using social media to predict the outbreak of violence in the 2011 London Riots

This paper considers the problem of understanding the influences of narratives or stories on individual and group behavior. Narrative theory describes how stories help people make sense of the world, and is being used to explain behavior in domains such as security, health care, and consumer behavior. We are interested in using narrative theory to develop better predictions of behavior and have developed a multi-methodology approach to combine narrative influence with system dynamics modeling of group behavior. Our model quantifies how individuals use narratives to understand current events and make decisions. We model the time-varying strength of cultural narratives as a degree of belief in the narrative’s explanatory power, updated heuristically in response to observations about similarity between cultural narratives and current events. We use Twitter posts to measure narrative-significant observations in the real world. Using this approach, we investigate a case study of the violent riots in London in 2011 and demonstrate how relevant narratives can be identified, monitored, and included in behavior models to predict violent activity. Introduction The study of narratives is providing exciting new opportunities to better understand the role stories play in human psychology and sociology and help explain human behavior in many domains. These opportunities are especially notable in security contexts, where the inclusion of narrative impacts can help explain the factors that contribute to radicalization, violent social mobilization, and insurgency, among other challenges. For example, in 2007, the US House of Representatives Committee on Armed Services held a hearing to discuss the “Battle of Ideas” in the war on terrorism, and the task of “winning [our adversaries’] hearts and minds”. The hearing underscored the fact that often times violent extremism is driven by a specific underlying worldview and that it is important to understand the narratives which make up this worldview. (House Armed Services Committee, 2007) Just as conventional conflict requires an understanding of the contested physical terrain, a winning a battle of ideas will require understanding of the narrative landscape. . 1 Research Associate, Sloan School of Management, Massachusetts Institute of Technology 2 Principal Research Scientist, Sloan School of Management, Massachusetts Institute of Technology 3 Principal Consultant, PA Consulting Group This paper considers the problem of understanding the narratives of inter-group conflict. It develops a mixed methodology approach that combines system dynamics models with social media data in order to better explain and predict outcomes of inter-group conflicts, particularly the escalation of group behavior towards violence. In addition to drawing upon previous system dynamics modeling research, our work incorporates emerging research on narrative networks that give insight on how stories exert powerful influences on human thoughts and behavior. The theory of ‘Narrative’ influence on collective behavior is gaining traction in social psychological circles as an alternate to methods assuming rational economic actors; and with the potential to “serve as a barometer for public views” and attitudes. (Monitor 360, 2012). Our work formalizes critical dynamics relating to the way stories exert powerful influences on group behavior in a system dynamics model and utilizes social media data for parameterization. We believe that incorporating narrative perspectives in formal simulation models, and drawing upon social data-sets, is an exciting new opportunity for system dynamics research to help develop responses to better predict and mitigate inter-group conflicts. The paper is organized as follows. The first section provides a brief overview of the social psychology research into narrative influence that provides the theoretical basis for the quantitative model. The second section gives a detailed formulation of the system dynamics model. Subsequently the paper discusses the use of social media streams as a proxy for important model components. The third section is a case study in which the methods developed in the paper are applied to understand an incident of rioting in London in 2011. Finally, the paper concludes with discussion of lessons learned from the case study and their implications for theory and practice. Brief overview of narrative theory Narrative theory, broadly defined, is concerned with how narratives influence cognition and behavior. Over the last several decades, researchers have investigated the origin, transformation, and behavioral impact of narratives. A central premise of narrative theory is that human beings understand their lives in terms of stories. While narratives can come from many sources, most of the narratives an individual maintains “derive from public narratives conveyed... by other people; things... intentionally taught” and come from a pool of cultural stories which cover “religion, politics, popular culture, regional identity, racial and ethnic identity, attitudes toward other members of the culture and toward minority members, attitudes toward outsiders” (Beach, 2010, p. 30). Narratives have an important role in consolidating memories, shaping emotions, and providing group distinctions, among other impacts. Other authors have described the impact of narrative on behavior, by arguing that: “humans use... narratives to understand the world and their place in it, to frame events, and to plan and justify their actions” (Halverson, Goodall, & Corman, 2011, p. 181) , and that narratives are how humans learn “ what is right and wrong and why, how and why things happen, how to perform tasks to produce desired effects... who we are, where we fit in the scheme of things, and what our rights and responsibilities are.” (Beach, 2010, p. 29) A well-cited phrase that describes the role narratives play in behavior—“I can only answer the question ‘What am I to do?’ if I can answer the prior question ‘Of what story or stories do I find myself a part?’” (MacIntyre, 1981)—helps capture our central interest in using narratives to better understand human behavior. Identifying relevant narratives The first step in utilizing narrative theory in our simulation models is to identify the narratives relevant for a target population If we can understand the narratives present in a population, and measure their shifting influences on that population’s attitudes and decision making, we gain insight into that population’s potential collective behavior and can prepare an informed response. The task of identifying narratives present in a population has been addressed by several sociologists in a limited number of case studies. For example: researchers have investigated narratives of Islamic Extremism (Halverson, Goodall, & Corman, 2011); narratives relevant to the Palestinian/Israeli Conflict (Hammack, 2009); narratives related to the 2012 US presidential election (Rosenstiel & Jurkowitz, 2012); and narratives with influence on the behavior of spouses in long distance relationships (Bergen, 2010). There are many narrative researchers expanding the range of population narratives, and we anticipate that as research progresses there will be increasingly better mappings between populations and narratives. We believe that investigating historical precedent, news articles, cultural literature, and social media can serve as input for us to achieve a rough understanding of the narratives related to a problem of interest. Building a quantitative model of narrative influence Our goal is to increase the predictive capacity of narrative theory, and we therefore need to move beyond only mapping narratives to populations. Our goal is to establish a quantitative method to determine the strength and influence of narratives, and ultimately their impact on behavior. This section describes the key pieces of the model and our approach to model development that will allow us to start evaluating real world cases. A metric of narrative strength Cultures maintain multiple narratives in order to “assure a closer correspondence between the narrative and the underlying facts in any given case.” (Skeel, 2009, p. 1203) Narratives with the most fidelity then have the capacity to direct behavior. Fisher explains that narratives which “represent accurate assertions about social reality... constitute good reasons for belief or action.” (Fisher, 1987, p. 105) Narrative strength refers to power of each narrative as they compete among multiple narratives to explain the world as it is observed— the strongest narrative will win out in this competition. We model narrative strength as the probability, that an individual sees themselves as participants in a particular cultural narrative. Narrative building and decay Our next step is to model how the strength of each narrative is built over time. According to Dehghani et al, “the rate of retrieval of cultural narratives [is dependent] upon the degree of surface and structural similarity with the presented [situation]” (Dehghani, Sachdeva, Ekhtiari, Gentner, & Forbus, 2009). In this manner, individuals update their beliefs about narratives by assessing the similarity of current events with their cultural narrative, with greater fidelity leading to greater narrative strength, as seen in Figure 1. Figure 1: Narratives compete to influence behavior. To simplify our model, we start with the assumption that each observation of narrative fidelity contributes equally to the strength of that narrative. This simplification is appropriate at the level of the population with the joint assumption that more influential events are more readily observed. Our focus on ‘observations’ of events captures much of this effect. This formulation is in line with previous work that models the role of messaging in the growth of insurgency. (Choucri, Goldsmith, Madnick, Morrison, & Siegel, 2007) Next, we focus on the strength of narrative strength over time. Consistent with system dynamics modeling of decision making, we make the assumption that new observations are more relevant to an individual’s assessment of narrative fidelity than are older ones, and as such have more influence over narrative strength.(Sterman, 2000) Similarly, as an observation of narrative fidelity is succeeded by others, its influence over the individual’s assessment of their situation declines. We next formulate an approach to measure the relative strength of a new observation versus an older observation. We model the strength of each observation as it decays with the addition of subsequent observations. Figure 2 illustrates this formulation. This formulation is consistent with prior modeling work on advertising. (Brady, 2009) Figure 2: The influence of observations B upon narrative strength of A as it decays with subsequent message volume (Hypothetical Observations) As more and more observations are made, the strength of the oldest observations approach zero as they become less and less relevant to the present situation. To formalize this relation, we assume that the ‘strength’ of each observation decays exponentially with the number of subsequent observations. If the initial strength of an observation is ‘a’, and ‘n’ subsequent observations have been made, then we can describe the residual strength of an individual observation as: Strengthobservation = a ∙ (1 − a)n The strength of a narrative (say ‘Narrative A’) is the sum of the residual strengths of all the observations which contribute to that narrative: StrengthA = �a ∙ (1 − a)ni i We can model this by tracking a level for each narrative, and when a new observation is made, multiplying all narrative levels by (1-a) before adding the new observation’s strength to the relevant narrative. Say that observation ‘i’ contributes to Narrative A, then upon its arrival, all narratives would update as: StrengthA[i] = StrengthA[i − 1] ∙ (1 − a) + a StrengthB[i] = StrengthB[i − 1] ∙ (1 − a) ... We need to re-map this to the time domain for inclusion in a System Dynamics model. In the time domain, especially when we discretize our models, we may have multiple observations occurring in the same time period, and contributing to different narratives. We lose some information about the order in which observations are made, and so we need to add their contributions to the strengths of each narrative in a way that does not preference one over the other without cause. If we knew the order that ‘m’ messages arrived, we would know that at the end of the time period, those messages would have contributed a certain fraction of total strength of all narratives equal to: 0 0.25 0.5 0.75 1 0 5 10 15 20 25 30 35 40 45 50 55 60 Strength decays with subsequent observations Narrative B Strength Narrative A strength Observations