We aim to quickly and accurately find hidden patterns in large graphs (i.e., collections of nodes and edges) that are growing in time.
Graph data is widespread---from social to biological networks. Notonly is there an increasing amount of graph data, but the graphs
themselves are growing in time. As with other types of big data, we
are interested to capture complex structure in the data, quantify our
uncertainty about unknown quantities of interest, and encode any prior
information that we might have. In theory, "Bayesian nonparametrics"
provides a set of methods to satisfy these desiderata. In practice,
little work has been done so far on learning in graph data, which
poses a unique set of problems above and beyond more traditional
vector-valued data points. Our current work focuses on the
development of Bayesian nonparametric methodology for learning in
network data. Using the flexible tools of Poisson point processes and
completely random measures, we aim to develop accurate probabilistic
models to capture graph behavior both at particular points in time as
well as across time as more graph data accrues. Using these models, we
aim to develop scalable algorithms for inference in networks.