Enhanced Attribute-Based Explanations of Multidimensional Projections

Abstract

Multidimensional projections (MPs) are established tools for exploring the structure of high-dimensional datasets to reveal groups of similar observations. For optimal usage, MPs can be augmented with mechanisms that explain what such points have in common that makes them similar. We extend the set of such explanatory instruments by two new techniques. First, we compute and encode the local dimensionality of the data in the projection, thereby showing areas where the MP can be well explained by a few latent variables. Secondly, we compute and display local attribute correlations, thereby helping the user to discover alternative explanations for the underlying phenomenon. We implement our explanatory tools using an image-based approach, which is efficient to compute, scales well visually for large and dense MP scatterplots, and can handle any projection technique. We demonstrate our approach using several datasets.