Graphs, Cuts and p-Spectral Clustering
Event details
| Date | 04.05.2009 |
| Hour | 16:15 |
| Speaker | Dr. Matthias Hein, Saarland University, Germany |
| Location |
INM202
|
| Category | Conferences - Seminars |
Graph-based methods can be applied to any kind of data and are
therefore heavily used in practice. In machine learning most of
the time so called similarity graphs are used. Although the
practical results of learning algorithms depend heavily on the
graph construction, this is a largely unexplored area in machine
learning. In this talk I show results from manifold learning and
clustering which illustrate the influence of graph type and graph
parameters. In particular, I present recent results which show
that the population version of the clustering objective induced
by the normalized cut criterion depends on the employed graph type.
In the second half of the talk I discuss a generalized version of
spectral clustering based on eigenvectors of the graph
p-Laplacian, a non-linear generalization of the graph Laplacian.
Interestingly, one can prove that the cut value obtained by
thresholding the second eigenvector of the p-Laplacian converges
towards the optimal Cheeger cut as p tends to 1.
Bio: Matthias Hein has been researcher from 2002 to 2007 at the
Max-Planck-Institute for Biological Cybernetics in the Empirical
Inference group of Prof. Schoelkopf. He received his doctoral
degree in Computer Science in 2005 from the Technical University
Darmstadt. Since 2007 he is Juniorprofessor at the Computer
Science Department at Saarland University.
M. Hein's homepage
Practical information
- General public
- Free