Evaluating results of K-medioids with no label data

users
#1

Hi,

I’m running a clustering analysis based on different distance metrics (cosine, mahalanobis, euclidean) and I want to be able to evaluate my results with some metric (sillouette or other).

I can’t use the scorer node because I don’t have info about the real classification of my data, the clusterings are exploratory, anyone has some idea how to compute some measure that helps me evaluate the different clustering results??

Thanks

0 Likes
#2

Hi @Cpalomeque -

You might try a couple of different options:

  1. @aborg’s implementation of the Silhouette method in KNIME (see here for more info: Metrics Nodes for Cluster Analysis Missing ?)

  2. Calculation via an R Snippet or Python Snippet. The R cluster package in particular contains an implementation of the Silhouette method that is relatively straightforward (simple example here: https://stackoverflow.com/questions/33999224/silhouette-plot-in-r)

1 Like