Usually, my go-to goodness of fit for evaluating clustering (e.g., k-means) is the average silhouette.

However, for DBSCAN it doesn't work since there are lots of non-clustered points. So increasing the stringency of the parameters removes neighboring clusters, and therefore increases the difference between intra- and inter-cluster distance -- just because there are less points clustered.

What is a good clustering method for DBSCAN that takes into account this issue or is robust to number of points clusters?

  • $\begingroup$ how about treating each non-clustered point as a single cluster? $\endgroup$ – dontloo Jul 31 '18 at 4:05

The only measure that I know that takes noise points and density into account is DBCV:

Moulavi, D., Jaskowiak, P. A., Campello, R. J., Zimek, A., & Sander, J. (2014, April). Density-based clustering validation. In Proceedings of the 2014 SIAM International Conference on Data Mining (pp. 839-847). Society for Industrial and Applied Mathematics.

I haven't used it though.


@thc there is another paper worth looking,

Jaskowiak, P.A., 2015. On the evaluation of clustering results: measures, ensembles, and gene expression data analysis (Doctoral dissertation, Universidade de São Paulo).

It discusses in detail the validation of density based clustering methods including the DBCV suggested earlier. It also includes a rich discussion on related works and experimental evaluation.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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