normalised mutual information

Normalised mutual information is an information measure used to assess the quality of clustering and classification algorithms (concept learning) in general. If you split classified data into several groups, then the entropy of the result will be higher, even if the division is entirely random and unconnected to the target classification, simply becasue the reulting groups and hence probabilities are smaller. It is obtained by dividing the raw information gain (gain(C,G)) by the sum of the entropy of the classification and decision grouping. That is
  Normalised information gain   =   gain(C,G) / ( H(C) + H(G) )
where H(G) is the entropy of the division into the groups.

. This is similar to but different from normalised information gain, as the latter just divides by H(G). In practice this means normalised mutual information it is weighted slightly less against larger numbers of groups compared with normalised information gain.