Thx Yuval

I already implemented a different way to create a cluster center and it
works great,
but I wanted to use the existing Mahout implementation and not develop my
own if possible...

I think it is not the Tanimoto distance measure that gives such a centroid
- I think it is the way a cluster center is created from points
(which currently doesnt depend at all in the distance measure)

It creates a center "mathematically legal" for kmeans
on Real numbers metrics
but doesnt have much sense in my my problem
which is a discrete problem

i.e. it will not cluster users into clusters by having them read the same
documents well enough

[Im not interested in the centroid (except for visualization and analysis
which is also important)  - it is just that the clusters created seems
logically wrong to me, and most users were clustered into 1 or 2 clusters
with many-many features - too many - with low weights. My new
implementation didnt suffer from this problem]

I will think on this some more

Thx all for your help on this issue




On Thu, Jun 21, 2012 at 4:32 PM, Yuval Feinstein <[email protected]>wrote:

> Shlomy, I will try to paraphrase what I understood:
> a. You are interested in the centroid for its own sake, rather than the
> cluster.
> b. An ideal centroid will have only zeros or ones as its coordinates, and
> thus create a summary of the cluster.
> c. The current Tanimoto implementation gives you a different centroid, with
> many small coordinate values.
>
> I do not know enough of the math for K-means, but you could implement your
> own distance function,
> that will somehow penalize non-integral coordinates/distances.
> I am not sure if such a function will generate a converging K-means
> algorithm,
> but it might be worth a shot.
>
> If you are looking for summaries of vector sets, you can try matrix
> factorizations
> (PCA,SVD,NNMF), and see which dominant components come out.
>
> HTH,
> Yuval
>
> On Thu, Jun 21, 2012 at 2:43 PM, Sean Owen <[email protected]> wrote:
>
> > Erm, I think I am thinking of canopy clustering.
> >
> > For k-means, I suppose you could say the choice of the k means isn't
> > quite right if k-1 of them are nowhere near most points. I don't know
> > how they were chosen. But again maybe not the real issue at heart
> > here.
> >
> > On Thu, Jun 21, 2012 at 12:36 PM, Shlomy Boshy <[email protected]>
> > wrote:
> > > Interesting - to which thresholds do you refer?
> >
>

Reply via email to