2
$\begingroup$

First of all, I am doing clustering and I have the true labels for my data. For evaluation, I am using the weighted average of the entropy values for each predicted cluster. I also came across with Mutual Information as a similar approach while going over the alternatives. On my data, they seem to give similar results.

However there is one issue that puzzles me.

Given the predicted cluster set $U$ and true clusters $V$, mutual information was defined as: $$ I(U,V) = H(U) - H(U|V) $$ or, $$ I(U,V) = H(V) - H(V|U) $$ If my math is correct, the average entropy that I'm using corresponds to conditional entropy term $H(V|U)$ and trying to minimize this aligns with maximizing the mutual information.

What I cannot see is how weigthed average entropy would differ from mutual information and why we would need the entropy terms $H(U)$ or $H(V)$. It feels like minimising one of the conditional entropies should suffice.

To put it another way, as far as I understood from the equations, having high entropy for true or predicted clusters in itself also results in higher mutual information. Does this mean that mutual information favors equally-sized clusters?

Thanks in advance.

$\endgroup$
1
$\begingroup$

Mutual information does favor many small clusters. Nectar these tend to be "pure". That is why variations wish as normalized mutual information and adjusted mutual information (AMI) are used instead.

Vinh, N. X.; Epps, J.; Bailey, J. (2009). "Information theoretic measures for clusterings comparison". Proceedings of the 26th Annual International Conference on Machine Learning - ICML '09. p. 1. doi:10.1145/1553374.1553511.

$\endgroup$
  • $\begingroup$ Thanks for the reference! It also suggests normalized information distance (NID) as a general purpose measure. $\endgroup$ – Esmailian Apr 19 at 20:25
  • 1
    $\begingroup$ There was also some follow-up that argues for different adjustments for chance. $\endgroup$ – Anony-Mousse Apr 20 at 6:58
  • $\begingroup$ I did not take a close look on adjusted and normalized versions. So, they essentially prevent the metric from favoring any particular number of clusters. Thank you! $\endgroup$ – ismlhk Apr 22 at 7:09
  • 1
    $\begingroup$ They don't adjust for the number of clusters. They adjust for random permutation of labels with the same cluster sizes. AFAICT, NMI still favors many small clusters. $\endgroup$ – Anony-Mousse Apr 22 at 15:29

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.