| New technologies have made it possible to easily collect information about social networks as they are acted and observed "in the wild," instead of as they are reported after-the-fact in surveys. This unprecedented access to social behavior data---data that captures the observable actions of multiple people as they interact with one another---provides opportunities to address many new research questions: How does local behavior relate to the global structure of the social network? How does a social network change over time? How can meaningful information be extracted from raw, recordable data? And how can all of this be done while protecting privacy?;With the goal of answering those questions, this dissertation presents new methods for measuring and modeling social networks derived from automatically recorded behavioral data. These techniques are presented in three parts.;First, new methods that use privacy-sensitive audio data to automatically find colocated people, determine who is conversation with whom, and detect who speaks when and how (pitch, rate, etc.) are presented. The use of these methods to gather a data set capturing a year's worth (426 person-hours) of real-world face-to-face conversations within a subject population of 24 graduate students is then recounted.;Second, two new extensions to exponential random graph models are proposed. These extensions exploit the richness of social behavior data and enable the new models to: (i) recover latent networks where hidden social relationships are observable only through noisy behavior data, and (ii) discover long range, high level properties of evolving social networks using time-inhomogeneous models.;Third, an influence mixture model is proposed that quantifies the amount of influence each person in a multi-person interaction exerts on the all of others. This measured influence is found to correlate positively with a person's centrality in her social network. |
