Using Multiple Clustering Algorithms to Generate Constraint Rules and Create Consensus Clusters

Data clustering techniques is used for aiding knowledge discovery when no additional information is available. There are several clustering techniques which produce reasonable results, although they often produce qualitatively distinct clusterings. In this paper, we study how different clustering algorithms produce different kinds of clusters and their relations. Also, we evaluate the possibility to merge differently generated clustering into a new clustering which neither of original algorithms can produce. The main contribution of this paper is a new algorithm which merges previous generated clusterings based on must-link constraint rules built from agreements among elements observed from such clusterings. This novel approach employs the entropy of agreements in order to decide to which cluster should an element belong. Experimental results indicate: 1) our approach can merge characteristics from original clusterings; 2) in some situations, it captures new information from data and improve results, mainly when considering external perspective; and 3) in no situation it has produced significantly worse results.