Osom: A method for building overlapping topological maps

We know that overlapping clustering solutions extract data organizations that are more fitted to the input data than crisp clustering solutions. Moreover, unsupervised neural networks bring efficient solutions to visualize class structures. The goal of the present study is then to combine the advantages of both methodologies by the extension of the usual self-organizing maps (Som) to overlapping clustering. We show that overlapping-Som allow to solve problems that are recurrent in overlapping clustering: number of clusters, complexity of the algorithm and coherence of the overlaps.We present the algorithm Osom that uses both an overlapping variant of the k-means clustering algorithm and the well known Kohonen approach, in order to build overlapping topologic maps. The algorithm is discussed on a theoretical point of view (associated energy function, complexity, etc.) and experiments are conducted on real data.

► We present a new approach for overlapping clustering using topological maps.
► Overlapping topological maps solve usual limitations of overlapping clustering.
► The approach generalizes existing overlapping and topographic clustering models.
► We propose a general evaluation framework to show the benefits of the new method.
► Experiments conducted on real multi-label datasets show the relevance of the method.