In the experiments, since the objects in the training datasets have been labeled, the way of generating the constraint set is very simple: the points of must-link are randomly selected from the same clusters and the points of cannot-link are selected randomly from different clusters.