You are viewing a plain text version of this content. The canonical link for it is here.
Posted to user@mahout.apache.org by Mark <st...@gmail.com> on 2011/06/26 21:29:05 UTC
KMeans and Canopies
Should canopy generation and KMeans clustering typically use the same
distance calculation or is possible to mix and match? Any reason why
some would mix?
Thanks
Re: KMeans and Canopies
Posted by Mark <st...@gmail.com>.
What if one was to use Tanimoto distance measure with KMeans.. would the
same reasoning apply?
On 6/26/11 1:54 PM, Christoph Brücke wrote:
> Hi Mark,
>
> you typically choose a somewhat cheaper distance metric for the canopy clustering, if used as a preprocessing step for KMeans. A simple example would be Manhattan distance (d = |x1 - x2| + |y1 - x2|) for Canopy clustering and Squared Euclidean distance [d = sqrt( (x1 - x2) ^2 + (y1 - y2) ^ 2) )] for KMeans. This way you got a cheap approximation for your initial cluster centers.
> I hope this was helpful.
>
> Regard,
> Christoph
>
>
> Am 26.06.2011 um 21:29 schrieb Mark:
>
>> Should canopy generation and KMeans clustering typically use the same distance calculation or is possible to mix and match? Any reason why some would mix?
>>
>> Thanks
>>
> Christoph Brücke
> christoph.bruecke@campus.tu-berlin.de
>
>
>
Re: KMeans and Canopies
Posted by Christoph Brücke <ch...@campus.tu-berlin.de>.
Hi Mark,
you typically choose a somewhat cheaper distance metric for the canopy clustering, if used as a preprocessing step for KMeans. A simple example would be Manhattan distance (d = |x1 - x2| + |y1 - x2|) for Canopy clustering and Squared Euclidean distance [d = sqrt( (x1 - x2) ^2 + (y1 - y2) ^ 2) )] for KMeans. This way you got a cheap approximation for your initial cluster centers.
I hope this was helpful.
Regard,
Christoph
Am 26.06.2011 um 21:29 schrieb Mark:
> Should canopy generation and KMeans clustering typically use the same distance calculation or is possible to mix and match? Any reason why some would mix?
>
> Thanks
>
Christoph Brücke
christoph.bruecke@campus.tu-berlin.de