For manifold learning, we projected the high-dimensional dataset either to a graph (e.g., of k nearest neighbors KNN, and variants such as mutual nearest neighbors) or to two or three dimensions (using PCA, t-SNE, SFDP). We also combined these approaches, first projecting to a graph, then calculating distances on the graph (e.g., Jaccard distance, or multiscale KNN distance; see below), then using those distances to project to 2D space using graph-t-SNE (gt-SNE; see below).
