I'm currently trying to use cluster analysis as a tool for time-series aggregation for a project of mine. The dataset is high-dimensional (386-d), so no chance in assessing the cluster validity visually.

I'm using three different clustering algorithms (k-means++, k-medoids PAM, fuzzy c-means) to find representative periods. As I do not know how many periods (thus, how many centers/medoids k) are present in the dataset, I want to use an internal cluster validity index (cvi) for it. (Basic procedure: run the clustering with multiple k's and plot the cvi against the k; choose highest/lowest k dependent on cvi optimum).

Let's stick with k-means as an example. It is non-deterministic, thus I initiate it multiple times with varying starting points. It then tries to reduce the intra cluster variance. The result with the lowest intra cluster variance of the various initiations with the same k is then kept.

My question is: should the cluster validity index that I use consider the intra cluster variance as a cohesion measure? So to speak: should the cvi use the objective function of the clustering algorithm as a measure to assess "goodness" of clustering?

On the one hand I think, that this would be a good idea, as it can assess, where the clustering algorithm was "most successfull" in its objective function. On the other hand I think, that for a good clustering it shouldn't depend on the same (cohesion) measure. Thus, using the same measure would increase the likelihood of identifying a "bad" clustering, that might represent an unnatural cluster, drawn from random points.

What's "the truth" here?


1 Answer 1


It's standard to use, e.g., Silhouette to assess the quality of clusterings, that were obtained with other methods.

I'm not a big fan of this (people always think they would get the "optimal" result this way, but they don't), but it's all over literature. Logically, you are trying to maximizing Silhouette and use the other algorithms as sampling procedure for possible results.

  • $\begingroup$ Thank you for your answer! Could you explain what you mean with "use the other algorithms as sampling procedure for possible results"? Btw: I read and used silhouette score for assessment, but it keeps having its maximum at k=2, whereas Calinski-Harabasz and Dunn's Index show maximum at k=4, for example. So I figured, Silhouette might not be the "best" choice for my problem. And by "people always think they would get the 'optimal' result this way, but they don't" I believe you mean, that the CVI does not necessarily peak at the optimal k, such that the results should be viewed with caution? $\endgroup$
    – Lens
    May 29, 2019 at 9:04
  • $\begingroup$ Yes, the indexes peak at some value, but that is not necessarily the most useful k, maybe not even close. Plus, there is always some peak even if all the results are bad because the data was not prepared well or is not suitable for this clustering... $\endgroup$ May 29, 2019 at 21:37
  • $\begingroup$ Yes, I am aware of this. But almost certainly, silhouette value is not suitable for every application, otherwise there wouldn't be so many options. Therefore, my question is, if it is helpful to use the same criteria for cohesion/separation measurement in the CVI as in the clustering algorithm. $\endgroup$
    – Lens
    May 31, 2019 at 7:57
  • $\begingroup$ Don't choose the CVI (or it's parameters) because of the clustering. Choose both the clustering and the CVI because of the problem that you are trying to solve. $\endgroup$ May 31, 2019 at 8:17

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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