Thesis

Social Evolution: Modeling Opinions and Behaviors in Face-to-Face Networks

Madan, A. "Social Evolution: Modeling Opinions and Behaviors in Face-to-Face Networks"

Abstract

Exposure to new ideas and opinions, and their diffusion within social networks, are im- portant questions in education, business, and government. However until recently there has been no method to automatically capture fine-grained face-to-face interactions between people, to better model the diffusion process. In this thesis, we describe the use of co- location and communication sensors in `socially aware' mobile phones to model the spread of opinions and behaviors of 78 residents of an undergraduate residence hall for an entire academic year, based on over 320,000 hours of behavior data.

Political scientists (Huckfeldt and Sprague, APSR, 1983) have noted the problem of mutual causation between face-to-face networks and political opinions. During the last three months of the 2008 US presidential campaigns of Barack Obama and John McCain, we find that political discussants have characteristic interaction patterns that can be used to recover the self-reported `political discussant' ties within the community. Automatically measured mobile phone features allow us to estimate exposure to different types of opinions in this community. We propose a measure of `dynamic homophily' which reveals surpris- ing short-term, population-wide behavior changes around external political events such as election debates and Election Day. To our knowledge, this is the first time such dynamic homophily effects have been measured. We find that social exposure to peers in the network predicts individual future opinions (R2 = 0:8; p < 0:001). The use of mobile phone based dynamic exposure increases the explained variance for future political opinions by up to 30%.

It is well known that face-to-face networks are the main vehicle for airborne contagious dis- eases (Elliott, Spatial Epidemiology, 2000). However, epidemiologists have not had access to tools to quantitatively measure the likelihood of contagion, as a function of contact/exposure with infected individuals, in realistic scenarios (Musher, NEJM, 2003), since it requires data about both symptoms and social interactions between individuals. We use of co-location and communication sensors to understand the role of face-to-face interactions in the contagion process. We find that there are characteristic changes in behavior when individuals become sick, reflected in features like total communication, temporal structure in communication (e.g., late nights and weekends), interaction diversity, and movement entropy (both within and outside the university). These behavior variations can be used to infer the likelihood of an individual being symptomatic, based on their network interactions alone, without the use of health-reports. We use a recently-developed signal processing approach (Nolte, Nature, 2008) to better understand the temporal information flux between physical symptoms (i.e., common colds, influenza), measured behavior variations and mental health symptoms (i.e., stress and early depression).

Longitudinal studies indicate that health-related behaviors from obesity (Christakis and Fowler, 2007) to happiness (Fowler and Christakis, 2008) may spread through social ties. The effects of social networks and social support on physical health are well-documented (Berkman, 1994; Marmot and Wilkinson, 2006). However, these studies do not quantify actual face-to-face interactions that lead to the adoption of health-related behaviors. We study the variations in BMI, weight (in lbs), unhealthy eating habits, diet and exercise, and find that social exposure measured using mobile phones is a better predictor of BMI change over a semester, than self-report data, in stark contrast to previous work.

From a smaller pilot study of social exposure in face-to-face networks and the propagation of viral music, we find that phone communication and location features predict the sharing of music between people, and also identify social ties that are `close friends' or `casual acquaintances'. These interaction and music sharing features can be used to model latent influences between participants in the music sharing process.

Related Content