Abstract: In this paper,a new clustering method is presented. According to this method, an initial class is found first, then set up a model of bipa-rtite graph relevant to sample points and feature changing intervals, and divide the primary class further with the bipartite model.This procedure is repeated until one gets the reasonable classes for all points. This method can be implemented by the calculation of the relational data base, and it is not necessary to fix numbers of classes in advance.We have got a good result from trial calculation of an example for earthquake forecast.