A scalable generative topographic mapping for sparse data sequences

We propose a novel, computationally efficient generative topographic model for inferring low dimensional representations of high dimensional data sets, designed to exploit data sparseness. The associated parameter estimation algorithm scales linearly with the number of nonzero entries in the observations while still learning a truly nonlinear generative mapping. The latent variables of the model lie in a 2D space that can be used for visualisation. We discuss related work and we provide experimental results on text based documents visualisation as well as the exploratory analysis of Web navigation sequences.