[Dmbu-l] Data Seminar
harshal at bu.edu
Tue Nov 29 12:07:51 EST 2016
*Kernel methods for label transfer between graphs*
*Mark Crovella, Boston University*
*Friday, December 2, 2016 at 10:30am in Hariri Seminar Room*
*Speaker: *Mark Crovella, Boston University
*Title:* Kernel methods for label transfer between graphs
*Abstract:* Imagine you have two graphs. In one graph, some of the
vertices have labels, which could be categories like “Democrat”,
“Republican”, etc. The other graph is not labelled. However, there is a
small set of “landmarks” — pairs of nodes, one from each graph, that are
known to correspond in some manner. The problem I’ll discuss is how to use
the structure of the graphs, plus the correspondence between the small set
of landmarks, to predict labels for vertices in the unlabeled graph.
Applications include deanonymization of social networks and function
prediction in biological networks.
To address this question, first I will review two different arguments that
suggest that a natural tool for label transfer in a single graph is the
Regularized Graph Laplacian. However, it is not obvious how to extend the
Regularized Laplacian to apply to multiple graphs simultaneously. I will
then show that by using the kernel property of the Regularized Laplacian,
there is a nature role for landmarks to form a geometric “bridge” between
the two graphs. I will then show results indicating that this method is
quite effective in practice for the function prediction problem.
This is work in progress with Ben Hescott (Tufts), Tim Lim (BU), and Thomas
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the Dmbu-l