> I'd like to know how to take an existing clustering and without
> changing the clusters themselves I would like to project the cluster
> centerpoints onto lower dimensions (3 or 2 ideally) to help visualize
> the size and relative proximity of clusters chosen.  Is there a way to
> set the dimensionality used in the random linear projection without
> altering the clusters chosen?  Or alternatively, is there a way to
> load an existing vector file and reproject to lower dimensions?

You can load a vector file and reproject it to lower dimensions with
SimPoint. Just re-run SimPoint with "-k 1 -iters 1" to effectively
disable clustering, and tell SimPoint to project to 2 or 3 dimensions
with the "-dim" option, and tell it to save the projected vectors with
the "-saveVectorsTxtFmt" option. You can use these low-dimensionality
reprojected vectors with the cluster assignments from your initial
SimPoint run to visualize how vectors are assigned to clusters.

This technique can't be used to reproject the cluster centers, since
the cluster centers in the original clustering run used a different
projection matrix, and cluster centers are defined in the projected
space. It should be possible [although I haven't tried] to save the
projected vectors and centers from the original clustering run, and
then use the technique above to project the already-projected vectors
and centers to 2 or 3 dimensions.

So with the first idea, the original vectors are being projected
twice, once to a higher dimensionality for clustering, then to a lower
dimensionality for visualization. With the second idea, the original
vectors are projected once to a higher dimensionality for clustering,
then the resulting projected vectors are projected again for
visualization.

-jeremy

---

YAHOO! GROUPS LINKS

•  Visit your group "simpoint" on the web.
   
•  To unsubscribe from this group, send an email to:
    simpoint-unsubscribe@yahoogroups.com
   
•  Your use of Yahoo! Groups is subject to the Yahoo! Terms of Service .

---