"Metaclustering""Metaclustering"http://www.haverford.edu/calendar/details/255301KINSC Hilles 109 2014-04-23T16:30:002014-04-23T18:00:00
April 23, 4:30PM–6:00PM
KINSC Hilles 109
Distinguished Visitor, Suresh Venkatasubramanian, University of Utah
Clustering is one of the most popular exploratory analysis tools available. Because it's unsupervised, it's easy to run without large amounts of training data. But because it's unsupervised, it's much harder to determine whether the resulting answers are good, or meaningful. Metaclustering is the larger effort of trying to take the outputs provided by clustering algorithms and derive answers that are meaningful, either by identifying patterns that are common among different methods, finding clusterings that are of good quality but give different perspectives on data, or even understanding how "robust" a particular grouping of data is. Doing this requires understanding the "shape" of the space of clusterings of data, as well as tools to manipulate this space. I'll discuss our recent efforts in this regard, as well as new work on determining how reliable a given clustering actually is. This is part of a larger focus on the problem of "accountability" in data mining.
Suresh Venkatasubramanian is an associate professor in the School of Computing at the University of Utah and is currently a visiting scientist at Google, Inc. His interests include algorithms, computational geometry, data analysis, and the challenges of large data problems. He is a recipient of a Warnock Endowed Chair at the University of Utah, as well as the NSF CAREER award. He also writes the Geomblog, a blog on algorithms, computational geometry and data.
Tea at 4:15PM
For More Info