Multiscale geometric feature extraction for high-dimensional and non-Euclidean data with application

Abstract: A method for extracting multiscale geometric features from a data cloud is proposed and analyzed. The basic idea is to map each pair of data points into a real-valued feature function defined on $[0,1]$. The construction of these feature functions is heavily based on geometric considerations, which has the benefits of enhancing interpretability. Further statistical analysis is then based on the collection of the feature functions. The potential of the method is illustrated by different applications, including classification of high-dimensional and non-Euclidean data. For continuous data in Euclidean space, our feature functions contain information about the underlying density at a given base point (small scale features), and also about the depth of the base point (large scale feature). As shown by our theoretical investigations, the method combats the curse of dimensionality, and also shows some adaptiveness towards sparsity. Connections to other concepts, such as random set theory, localized depth measures and nonlinear multidimensional scaling, are also explored.