Table 1.

Interpretations of commonly observed combinations of LA and NMI scores

NMI(A,B)	NMI(B,A)	LA	Implies
Low	Low	Low	Poor similarity
Low	High	Low	B refines A
High	Low	Low	A refines B
High	High	High	Good similarity

Given two clustering results A and B, for which both NMI(A,B), NMI(B,A) and LA(A,B) values are high (nearing the maximum value of 1.0), the two clusterings are very similar, and when all three are significantly lower, they are very different. But when NMI(A,B) is high, NMI(B,A) is low and LA is low, then it is likely that A is a refinement of B. In this case, many clusters in B have been broken into two or more clusters in A (possible combinations summarized in here) (1). The magnitude of dissimilarity that is important is defined by the user and may vary considerably with the dataset, although values <0.7 for both LA and NMI are usually viewed as quite different. Additional interpretation of differences measured by LA and NMI depends on more detailed analysis of the dissimilarities and their distribution over the dataset, as outlined above.