Spectral clustering

[Pfaendner]

Hey guys,

My Question is concerned to the spectral clustering variant 1 introduced in the lecture:

What is meant by 'first k eigenvectors'?

Are these the eigenvectors corresponding to the smallest k eigenvalues, or to the largest k eigenvalues, or even to the smallest non-zero eigenvalues?

Appriciate your help!

Greets Christian

[misbah baig]

Hi,

From the web-page's course information sections it says that exam will be held in two halls (Günter Hotz Hörsaal E 2.2 0.01 and E 2.5 HS I). Does anyone know about the sitting plan, is it published or still not announced ?

Regards,

Mirza

--
You received this message because you are subscribed to the Google Groups "Machine Learning WS18/19" group.
To unsubscribe from this group and stop receiving emails from it, send an email to machine-learning-...@googlegroups.com.
To post to this group, send email to machine-lea...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/machine-learning-ws1819/e4de8a07-6a3e-48ec-a6c7-163055a635f7%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Max Losch]

Hi,

First k eigenvectors refers here to the eigenvectors with smallest non-zero eigenvalues. This can be motivated by interpreting the graph as a system of nodes connected by springs with varying stiffness proportional to their pairwise similarities. To partition such a system, one is interested in looking at connections with low vibrating frequencies. These frequencies are determined via the eigenvalues of the graph laplacian.

For more in depth information see this lecture on spectral clustering:
https://people.eecs.berkeley.edu/~demmel/cs267/lecture20/lecture20.html

Best,

Max

[Pfaendner]

Thanks! That helps a lot for my understanding!