One of the problems arising in cluster analysis is a problem of comparing several different cluster partitions. The cause of such problems may-lie in using more than one method of clusters building or trying the same method with different starting configurations. In the paper we suggest an approach to comparing two cluster partitions of the same finite set. It is based on some metric on the class of all cluster partitions. A notion of stability for cluster algorithms is also defined and discussed.

Key words: cluster partition, stability, metric on finite set.