Samuel Kaski and Jaakko Peltonen

D imensionality reduction is one of the basic operations in the toolbox of data analysts and designers of ma chine learning and pattern recognition systems. Given a large set of measured variables but few observations, an obvious idea is to reduce the degrees of freedom in the measurements by representing them with a smaller set of more “condensed” variables. Another reason for reducing the dimensionality is to reduce computational load in further processing. A third reason is visualization. “Looking at the data” is a central ingredient of exploratory data analysis, the first stage of data analysis where the goal is to make sense of the data before proceeding with more goal-directed modeling and analyses. It has turned out that although these different tasks seem alike, their solution requires different tools. In this article, we show that dimensionality reduction for data visualization can be represented as an information retrieval task, where the quality of visualization can be measured by precision and recall measures and their smoothed extensions. Furthermore, we show that visualization can be optimized to directly maximize the quality for any desired tradeoff between precision and recall, yielding very well-performing visualization methods.