Generalization of c-means for identifying non-disjoint clusters with overlap regulation

• Two overlap regulation principles leading to new (parameterizable) overlapping.
• Discussion of the contribution w.r.t. the state-of-art on overlapping clustering.
• New evaluation approach for overlapping clustering evaluations.
• Experiments on real multi-label datasets show the efficiency of the contribution.
• Dispersal-based regulation principle builds reliable overlaps with an easy tuning.

Clustering is an unsupervised learning method that enables to fit structures in unlabeled data sets. Detecting overlapping structures is a specific challenge involving its own theoretical issues but offering relevant solutions for many application domains. This paper presents generalizations of the c-means algorithm allowing the parametrization of the overlap sizes. Two regulation principles are introduced, that aim to control the overlap shapes and sizes as regard to the number and the dispersal of the cluster concerned. The experiments performed on real world datasets show the efficiency of the proposed principles and especially the ability of the second one to build reliable overlaps with an easy tuning and whatever the requirement on the number of clusters.