Online social networks now provide a view of social interactions that is unmatched in scale, granularity and - equally important - amenability to automated analysis. However, only a small fraction of our social capital is spent online; moreover, online networks are typically only a reflection of richer, causative networks that govern our everyday interactions - relationships typically first develop offline. The aim of this proposal is to bring the full power of automated data-driven understanding to bear on the - arguably more important - social networks in the real world. The research will develop a scalable sensor-tag based infrastructure that measures co-locations of participants by dynamically pairing participants with tags and extract genuine interactions from mere colocations via theory of structure learning in Markov Random Fields. The research will characterize network properties (like subgroups, clustering and small worlds), and node properties (like centrality and influence) and capture the natural evolution of participant interactions via the theory of compressed sensing with sequential observations.
The techniques pioneered in this proposal will significantly advance our ability to obtain a meaningful data-driven understanding of social networks in the real world. Industry interaction will inform this research from the beginning, via established industry partnerships at UT Austin. Student involvement, both undergraduate and graduate, lies at the core of this research, providing a natural venue to introduce under-represented groups to networking research. Social networks provide a naturally engaging subject to introduce high-school students to engineering and networks, which the PIs will do via talks in area schools.
This project played a crucial role on several fronts: (a) It sponsored the development of new methods to learn social network structure from data on user actions; this is a the basic first step that needs to happen in the study of social network settings where the structure is not explicit. (b) It led to the development of new and state-of-art algorithms for gleaning structural properties that may be hidden in large graphs; for example, small communities of more-densely-connected nodes in large networks. (c) It led to new methods for inverse problems for epidemic processes. These processes are flexible models for a wide range of network pheomena; the inverse problem setting is one where we observe the spread of the process and need to infer what the underlying network is. (d) It led to a new understanding of non-convex methods for matrix inference; specifically, alternating minimization, which is a widely used but poorly understood method for matrix factorization and low-rank approximation. The project, over its lifetime, has funded/part funded 4 graduate students (three of whom have graduated with PhDs and landed jobs in academia). It has led to several publications in highly selective veues: both conferences like SIGMETRICS, NIPS, ICML, and journals like the IEEE transactions on Information Theory. It has also supported the development of new classes in stochastic processes, networks and optimization at UT Austin.
