Is there any studies/research on the correlation between extrinsic (Entropy, purity, B³ F1, ARI, etc.) and intrinsic (Silhouette, Calinski and Harabasz, Davies-Bouldin, etc.) clustering evaluation metrics?

The main reason for asking is assessing whether a certain representation of data is well suited to the problem. If a correlation holds, then analysing how the true clustering with the said representation fares, in terms of an intrinsic metric, could bound the expected clustering performance of a clustering algorithm in terms of an extrinsic metric.


Not necessarily.

Consider data with random labels. I'd be surprised to see a correlation of intrinsic and extrinsic measures. :-)

And from my experience with labels: no they are not well correlated with some cluster structure that you can find clearly.


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.