Mode Explorer: Using Model-based Sonification to Investigate Basins of Attraction

Abstract

This paper presents a novel interactive auditory data exploration method to investigate features of high-dimensional data distributions. The Mode Explorer couples a scratching-interaction on a 2D scatter plot of high-dimensional data to real-time dynamical processes, excited in data space at the nearest mode in the probability density function (pdf) obtained by kernel-density estimation. Specifically, the sign-inverted pdf is used as a potential function in which test particles perform oscillations at low friction, yielding signals that can directly be played back as sound. This Model-based sonification approach is used to interactively search the distribution for different modes, learn about their details, i.e. the Hessian matrix at the mode, and thus enable a non-parametric parameter selection for appropriate bandwidth.