On 03/25/2012 11:47 PM, Gael Varoquaux wrote:
> On Sun, Mar 25, 2012 at 11:38:50PM +0200, Andreas wrote:
>    
>>> Unlike something like spectral clustering, it is the euclidean distance
>>> to the centers that is minimized. Thus K-Means will seek clusters that
>>> are regular in the flat euclidean space.
>>>        
>    
>> Ok, that's right. Though I would argue that the distance measure
>> is not the only factor here. MeanShift also works with a euclidean
>> distance by using a Gaussian weighted density.
>> Here the difference is more in the objective then in the underlying
>> distance.
>>      
> Granted. Do you have suggestions for a better formulation? It would be
> helpful.
>
>    
As far as I can see, your groups are "KMeans + Ward" and "rest".
I don't know how ward works but looking at the lena example,
the clusters don't seem to be convex.
I'll try to understand the algorithm before I'll suggest anything
- maybe I should have done that before criticising your wording ;)

>
>>> But if you do it right, the two approaches solve the same problem, AFAIK.
>>>        
>    
>> They lead to different generalized eigenvalue problems:
>> http://arxiv.org/pdf/0711.0189
>>      
> Yes (that's exactly the reference that I had in mind), but they should
> minimize the same energy, right?
>
>    
Skimming the tutorial, I am not sure what the answer to that question is.
I think they are both relaxations of the normalized cuts problem,
but lead to different solutions in gerneral.

The tutorial suggests using Shi/Malik but since both are in
use in the literature, I thought it would be nice to have them both.

------------------------------------------------------------------------------
This SF email is sponsosred by:
Try Windows Azure free for 90 days Click Here 
http://p.sf.net/sfu/sfd2d-msazure
_______________________________________________
Scikit-learn-general mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/scikit-learn-general

Reply via email to