The Nystrom approximation for relational generative topographic mappings
Andrej Gisbrecht, Bassam Mokbel, Barbara Hammer · PUB – Publications at Bielefeld University (Bielefeld University) · 2010
Relational generative topographic mappings (RGTM) provide a statistically motivated data inspection and visualization tool for pairwise dissimilarities by fitting a constraint Gaussian mixture model to the data. Since it is based on pairwise dissimilarities of data, it scales quadratically with the number of training samples, making the method infeasible for large data sets. In this contribution, we transfer the Nystrom approximation to RGTM and we investigate its effect on the method. This leads to a linear method which reliability depends on the intrinsic dimensionality of the dissimilarity matrix.