Possible Duplicate:
How to tell if data is “clustered” enough for clustering algorithms to produce meaningful results?

I have used hierarchical clustering, e.g, Ward's method, single,complete, etc. I have the same problem, I do not know how to assess my clustering algorithm. How would I know that the clusters are meaningful?

I do not know what FPC software is, can any one use FPC, and, what is the programming language of this software?

If I'm not familiar with FPC, is there another means of assessment?

  • 2
    $\begingroup$ Can you write this question clearly enough for us to understand what you are asking? Maybe a data example would help. Otherwise the question should be closed as not a real question. $\endgroup$ Sep 10 '12 at 3:05
  • $\begingroup$ I guess "meaning" (semantics) is not a statistical category ... $\endgroup$
    – Felix S
    Sep 10 '12 at 6:00
  • 1
    $\begingroup$ Isn't this a duplicate of How to tell if data is “clustered” enough for clustering algorithms to produce meaningful results? $\endgroup$
    – chl
    Sep 10 '12 at 15:34
  • $\begingroup$ I've voted to close this in reference to the same question chl links to. If that prior q/a is not sufficient, please update this question for how your situation is unique. $\endgroup$
    – Andy W
    Sep 16 '12 at 15:17

Don't trust a mathematical measure. Analyze your data manually.

Otherwise, you risk "overfitting", in the sense that the clustering algorithm appears to be best that happens to be best correlated to your evaluation method. This happens all the time; people comparing k-means and other clusters based on the in-cluster variance. Now k-means tries to minimize in-cluster variance... so it is not surprising that it scores well on this measure. But you could say that this measure mostly tests how similar your clustering to a k-means clustering is.

Some of these measures have a well-defined use case. Assuming you run k-means multiple times (which is a best practice), which result should you choose? When comparing k-means with k-means, it is definitely okay to use such a measure. It will tell you which of the two the better k-means result is.

As for "fpc", it probably refers to the R clustering package "fpc". I believe it also includes some cluster validation indexes.


Not the answer you're looking for? Browse other questions tagged or ask your own question.