Orange Forum • View topic - Hamming, Relief, fitness

Hamming, Relief, fitness

A place to ask questions about methods in Orange and how they are used and other general support.

Hamming, Relief, fitness

Postby sapienza » Tue Sep 02, 2008 17:40

Hi there,

i would ask to you some explanations about hamming and relief: what do they are? i mean...i know how to calculate distances by manhattan and euclidean, but using them?

I would like even to know how it's done the fitness evaluator for k-means.

Any help is really appreciated


Postby Janez » Tue Sep 02, 2008 21:06


Hamming distance is well known (see For ReliefF, check Kononenko's papers. As I recall, it is the same thing as Manhattan distance with a few tricks for treatment of undefined values.


Postby sapienza » Wed Sep 03, 2008 10:16

Oh well really thanks. Very helpful ;-)

Relief seems different, uses probability:

W[A] = P (different value of A | nearest istance from different class) –
P (different value of A | nearest istance from same class)

By the way, i can't understand why BIC is calculated for every cluster: i thought it was an evaluator for the whole model. How can i obtain BIC for hierarchical clustering? I would like to check quality of k-means against hierarchical clustering.

Another thing is that at the end the values of BIC obtained for every cluster are added: does it mean that is an additive measure for every cluster, considered them as stand-alone models?

Last thing about BIC is: a good value of BIC is a low value in absolute value, or even considering the sign?

Thanks a lot for the help

Postby Janez » Wed Sep 03, 2008 10:45

The formula you wrote is what ReliefF uses for evaluation of attributes (in a similar fashion than, say, information gain of gini index). But this formula talks about the "nearest instances", and the part that you are interested in is how ReliefF determines which instances are "near".

As for BIC: I don't know. The answer is somewhere here:, especially here: ... ngCRS/src/.


Return to Questions & Support