Unfolding the Manifold in Generative Topographic Mapping

摘要: Generative Topographic Mapping (GTM) is a probabilistic latent variable model for multivariate data clustering and visualization. It tries to capture the relevant structure by defining low-dimensional manifold embedded in high-dimensional space. This requires assumption that can be faithfully represented of much lower dimension than observed Even when this holds, approximation may, some datasets, require plenty folding, resulting an entangled breaches topology preservation would hamper visualization cluster definition. partially avoided modifying GTM learning procedure so as penalize divergences between Euclidean distances from prototypes corresponding geodesic along manifold. We define assess strategy, comparing it performance standard GTM, using several artificial datasets.