Study of the Convergence in Automatic Generation of Instance Level Constraints

This work deepens in a methodology to generate Instance Level Constraints for Semi-supervised clustering by the study of the inherent nature of the data. The methodology executes a partitional clustering algorithm repetitively, so we study its behaviour according to the number of iterations of the clustering. In this scenario we propose three different stopping criteria to determine how many times the partitional clustering algorithm should be executed to obtain reliable instance level constraints. These criteria are experimentally tested under the document clustering problem.