Subspace Clustering Multi-module Self-organizing Maps with Two-Stage Learning

Clustering complexity increases with the number of categories and sub-categories and with data dimensionality. In this case, the distance metrics lose discrimination power with the growth of such dimensionality. Thus, we propose a multiple-module soft subspace clustering algorithm called Subspace Clustering Multi-Module Self-Organizing Maps (SC-MuSOM) that produces a map for each category. Moreover, SC-MuSOM learns a relevance coefficient for each dimension of each cluster handling the dimensionality curse. This fast-training model has a second learning stage in which the cluster prototypes are finely tuned considering the spatial resemblance between cluster centers. We validated the model with data mining sets from UCI Repository and computer vision data. Our experiments suggest that SC-MuSOM is competitive with other state-of-the-art models for the tested problems.