Measuring the quality of projections of high-dimensional labeled data

Dimensionality reduction techniques, also called projections, are one of the main tools for visualizing high-dimensional data. To compare such techniques, several quality metrics have been proposed. However, such metrics may not capture the visual separation among groups/classes of samples in a projection, i.e., having groups of similar (same label) points far from other (distinct label) groups of points. For this, we propose a pseudo-labeling mechanism to assess visual separation using the performance of a semi-supervised optimum-path forest classifier (OPFSemi), measured by Cohen’s Kappa. We argue that lower label propagation errors by OPFSemi in projections are related to higher data/visual separation. OPFSemi explores local and global information of data distribution when computing optimum connectivity between samples in a projection for label propagation. It is parameter-free, fast to compute, easy to implement, and generically handles any high-dimensional quantitative labeled dataset and projection technique. We compare our approach with four commonly used scalar metrics in the literature for 18 datasets and 39 projection techniques. Our results consistently show that our proposed metric consistently scores values in line with the perceived visual separation, surpassing existing projection-quality metrics in this respect.