Poster accepted at IEEE VAST 2018
Projecting data down to two dimensions to visualize in a scatterplot is one of the basic building blocks of visualization. While there are various established methods used for projection, many projections fail to capture phenomena at different scales, due to occlusion or overplotting. A trade-off emerges between showing small and large-scale structure. In this work, we present an algorithm that parameterizes this tradeoff to calculate multiple projections that vary by the scale of the highlighted structure. By jointly optimizing both the information theoretic content of the projection and the clipped bounding region of the resulting view, we can empirically find relevant structure to show to a user. We demonstrate the use of this algorithm on several synthetic and real datasets. We also describe how this method would be useful in a visual analytics system for providing a grand tour for both low- and high-dimensional datasets. By exposing a simple resolution parameter to the user, the user is able to guide their own path through their data, enabling them to glean multiple levels of insight in a way that other static projection techniques could not allow.