Improved data visualisation through nonlinear dissimilarity modelling

Abstract

Inherent to state-of-the-art dimension reduction algorithms is the assumption that global distances between observations are Euclidean, despite the potential for altogether non-Euclidean data manifolds. We demonstrate that a non-Euclidean manifold chart can be approximated by implementing a universal approximator over a dictionary of dissimilarity measures, building on recent developments in the field. This approach is transferable across domains such that observations can be vectors, distributions, graphs and time series for instance. Our novel dissimilarity learning method is illustrated with four standard visualisation datasets showing the benefits over the linear dissimilarity learning approach.